brand-setup
Onboard a new brand — configure API keys, connect integrations, analyze website, generate brand context files
SKILL.mdplugins/link-skills/skills/brand-setup/SKILL.md
GitHubMaintenance
| Agent | Version | Last Changed |
|---|---|---|
| Link | v2.20.0 | July 12, 2026 |
Description: Onboard a new brand — configure API keys, connect integrations, analyze website, generate brand context files
Change Log
v2.20.0 — July 12, 2026
- fb.ai API key is now project-scoped across 10 capabilities. Repointed all 5 key-generation links to the new API Keys page (
/dashboard/api-keys, reached from the profile menu; the old/dashboard/social-posts/api-keysstill 308-redirects) and added a scope-checkbox section at the point of generation: all 10 boxes are ticked by default and should be left that way;social_postsis the minimum for every skill shipping today; unticking a box makes those tools returnauth/insufficient-scope. Also documents the new key expiry (defaults to 90 days) and reframes the key — it is no longer "social templates", it unlocks SEO research, article generation, publishing, site audits, traffic monitoring and cold-email lead gen.
v2.19.0 — July 03, 2026
- Corrected Zernio ad-account discovery tool name (fixes the v2.18.0 migration bug). v2.18.0 assumed Zernio just drops the
late_prefix, but Zernio's tools are resource-prefixed. The Step 7 Step D / Step 8 ad-account probes now callads_list_ad_accounts(waslist_ad_accounts), and the two-ID docs referencead_campaigns_get_ads_timeline/ad_campaigns_list_ad_campaigns.profiles_list/accounts_listwere already correct.
v2.18.0 — July 01, 2026
- Zernio + DataforSEO split off the gateway (gateway v1.7.4 / v1.7.5). Both now ship their own MCP servers — they are no longer routed through the Five Agents gateway. Zernio is added as a required custom-connector (Step 2 row 2 + new Step 7a-ii,
https://mcp.zernio.com/mcp, OAuth) — the oldLATE_API_KEYpaste flow is gone. DataforSEO is added as an optional custom-connector (Step 2 row 7 + new Step 7a-iii,https://mcp.dataforseo.com/mcp, Basic Auth usingDATAFORSEO_LOGIN/DATAFORSEO_PASSWORD). Account discovery (Step 7 Step D) and the Step 8 validation probes now call the native tools (profiles_list,accounts_list,list_ad_accounts,keywords_data_google_ads_search_volume) with nofiveagents_api_keyparam. All${BRAND}_LATE_*env vars renamed to${BRAND}_ZERNIO_*(10 vars);LATE_API_KEYremoved from the key tables, save-list, and gateway vault. Existing brands are migrated byplugin-updatev2.18.0.
v2.17.0 — June 21, 2026
- Inbound Gig Engine connectors registered. Added n8n Cloud as an optional business-ops MCP (Step 2 overview row 15 + Step 7c walkthrough) — used by
n8n-workflow-builderto build the proof-of-concept workflow that backs an inbound gig bid; skip → that one skill is unconfigurable for the brand, the rest of the pipeline (gig-prospector → gig-proposal-writer) still runs. Added the optional${BRAND}_N8N_PROJECTenv var (Step 7b optional table) to target a specific n8n project/folder; absent → workflows land in the default project. Backfilled the missing${BRAND}_GIGS_DBrow (bootstrapped bygig-prospector, v2.16.0) into the auto-bootstrapped Notion DB table — no user action, created on first run like the others.
v2.16.0 — June 20, 2026
- New Step 5g Step H — Inbound Job Filters (powers the new
gig-prospectorSales skill). Captures the Markets to monitor (Singapore / Indonesia / Malaysia / Thailand / Australia / Global-Remote), the Platforms to scan — each tagged have account / no account — the Search Keywords (drafted fromproduct.md, i.e. what the brand actually sells — never hardcoded), and Budget Floor / Exclusions / Daily Cap. Written to a new## Inbound Job Filterssection in thesales.mdtemplate. Pre-fill mapping gainsproduct.md→ Step H (keywords) andbrand.md→ Step H (markets) rows. Skippable for brands that don't pursue marketplace work.
Brand Setup — New Client Onboarding
Before Executing
Read agents/link.md before starting. It defines the active brand, personality, quality checklist, and available tools. Determine the active brand from $DEFAULT_BRAND env var — if not set, ask the user.
You are the onboarding agent for the Link marketing plugin. Walk the user through setting up their brand step by step. Be friendly but efficient — explain what each step does and why.
When to invoke
- First time using the plugin
- Adding a new brand
- User says "set up", "onboard", "add brand", "configure"
Arguments
| Argument | Meaning |
|---|---|
-- project created |
User is already inside the project session. Skip step 1a and begin at step 1b. |
Flow
Run these steps in order. Do not skip or rush any step. At the end of each step, explicitly ask the user to confirm before proceeding to the next. Never assume a step is already done — always ask. The only exception is if the user explicitly says "skip" or "already done" for a specific step.
What this skill does
Brand setup creates the entire context library that the Link plugin's 23 skills read at runtime — the brand voice, target personas, product/pricing facts, competitive landscape, sales playbook, finance ops, and more. Once setup is complete, every other skill (/link-skills:research-strategy, /link-skills:content-creation, /link-skills:apollo-lead-prospector, etc.) can be run for this brand without re-asking the same questions.
Estimated time: 60–90 minutes for a typical SaaS brand. Can extend to 2 hours if you opt into all optional sub-steps (Claude Design system, social templates, fundraised-investor context). Most steps have smart defaults — you can usually answer "use the default" and keep moving.
The 10 steps:
| # | Step | Purpose | Time |
|---|---|---|---|
| 1 | Cowork Setup | Verify your Cowork project + permissions are configured correctly | 2–3 min |
| 2 | What You'll Need | Quick inventory of API keys and integrations you'll connect later | 1 min (review only) |
| 3 | Brand Name & Folder | Pick the brand slug, create brands/{brand}/ directory |
1 min |
| 4 | Website Analysis | Auto-extract tagline, voice, colors, personas from your website (Playwright + AI) | 5–10 min |
| 4b | Claude Design System (optional) | Install your Claude Design visual system for tighter brand consistency | 10 min if installed |
| 4c | Social Templates (optional) | Author Meta Carousel / Story / LinkedIn Post / Meta Post templates in Claude Design and upload them to the fb.ai dashboard (needs a paid fb.ai plan) | ~5 min per template |
| 5 | Research & Context Generation | Build out product.md, competitors.md, funnel.md + v2.4.0 sales/CS/finance/investors/operations files | 30–60 min |
| 6 | Logo | Copy your logo file into the brand folder | 1 min |
| 7 | API Keys & Connections | Connect each integration (FiveAgents gateway, Notion, Slack, Apollo, Stripe, Xero, etc.) | 15–30 min |
| 8 | Validate Connections | Test every connector with a low-cost call to confirm setup is good | 5 min |
| 9 | Initialize Workspace CLAUDE.md | Embed the Link agent into your workspace so all sessions auto-load it | 2 min |
| 10 | Summary & Completion Email | Final summary email + Slack DM with what was set up and what was skipped | 1 min |
You'll be asked to confirm at the end of each step before moving to the next — that's by design, so you can pause anywhere and resume later. The only exceptions are explicit "skip" responses.
Step 1 — Cowork Setup
Before anything else, ensure your Cowork environment is configured. Both 1a and 1b are mandatory — do not proceed until both are confirmed.
This step is purely about your Cowork environment — it doesn't create any brand files yet. Once you've confirmed your project + permissions, you'll be inside the project session for the rest of brand-setup.
1a. Work in a Project
All brand assets, outputs, and temp files live inside a Cowork project on your machine. This is required — brand setup cannot run without an active project.
Let's get your project set up. In Cowork:
- Look for Projects just below the chat input area
- Click "Create a new Project"
- Choose either "Start from scratch" or "Use an existing folder"
- Name your project (e.g. your brand name)
- Set Claude Permission as "Act without asking" — this ensures your scheduled jobs will run without needing your approval (after successfully tested). ⚠️ By enabling this, you agree to accept the risks raised by Claude when it acts autonomously. Review Claude's warnings carefully before confirming.
Once the project is created, ask the user to:
- Click the project name to open its session
- Inside that project session, run:
/brand-setup -- project created
Brand setup will resume from step 1b inside the correct project context. Do not continue in this session.
1b. Configure settings
To use this plugin, you'll need to adjust a few settings first. Go to Settings → Capabilities and check the following:
- Settings → Claude Code → Allow bypass permissions mode — toggle ON (required for skills to run without interruption)
- Settings → Capabilities → Domain Allowlist → All Domains — toggle ON (required for skills to fetch external URLs)
Have you enabled these settings? Once confirmed, we'll move on.
Do not proceed to Step 2 until the user confirms both 1a and 1b are done.
Step 2 — What You'll Need (Prerequisites Overview)
This step is just a heads-up — no questions yet, no setup actions. I'll walk you through everything you'll need to have ready before we get to Step 7 (API Keys). Skim it now so you know what to grab; we'll come back and ask for each item one at a time later. If you're missing anything, that's fine — most integrations are skippable, and you can always re-run brand-setup later to add them.
Before we begin, here's everything you'll want to have ready. You don't need all of these right now — we'll walk through each one — but having them handy will make setup faster.
Brand basics:
- Your brand name
- Your website URL (we'll auto-detect colors and fonts from your site)
- Your logo file path (PNG, transparent background preferred — e.g.
~/Documents/my-brand/logo.png) - A Claude account — we'll optionally create your visual design system at https://claude.ai/design (Step 4b — recommended for strongest brand consistency, but skippable; skills fall back to
brand.mdcolors/fonts when absent)
Required API keys:
| # | Key | What it's for | How to get it |
|---|---|---|---|
| 1 | FIVEAGENTS_API_KEY |
Dashboard logging, credential vault, email sending | 1. Go to https://fiveagents.io and sign in<br>2. Go to Dashboard → API Keys<br>3. Copy your fa_live_... key |
| 2 | GEMINI_API_KEY |
Image generation (social graphics, backgrounds) | 1. Go to https://aistudio.google.com/apikey<br>2. Click "Create API Key"<br>3. Copy the key (free tier: 10 images/min) |
| 3 | SLACK_NOTIFY_USER |
Slack DM notifications after each skill run | 1. Open Slack<br>2. Click your profile photo → "Profile"<br>3. Click the three dots ⋯ → "Copy member ID" |
| 4 | REPORT_EMAIL |
Daily/weekly marketing report delivery | Your work email address |
Required for social publishing: Zernio needs no API key — it's an OAuth custom connector (see the MCP connections table above, row 2, and Step 7a-ii). In short: sign up at https://zernio.com, create a Profile and connect your social accounts (Facebook, Instagram, LinkedIn, …) via OAuth in the Zernio dashboard, then in Claude go to Settings → Connectors → Add custom connector → https://mcp.zernio.com/mcp and sign in with your Zernio account.
Optional (SEO/GEO Research & Video Generation):
| # | Key | What it's for | How to get it |
|---|---|---|---|
| 5 | DATAFORSEO_LOGIN + DATAFORSEO_PASSWORD |
Keyword research (search volume, suggestions) — entered into the DataforSEO connector (Step 7a-iii) | 1. Sign up at https://dataforseo.com<br>2. Go to Dashboard → API Settings<br>3. Copy your login email and API password |
| 6 | FIVEBUCKS_API_KEY |
The brand's fb.ai content + traffic + lead-gen stack — paid fb.ai subscription (see Step 4b for what fb.ai is). Unlocks branded social templates (Step 4c), and also SEO research, article generation, publishing to 11 platforms, site audits, traffic monitoring, and cold-email lead gen | 1. Go to https://www.fivebucks.ai/dashboard/api-keys and sign in (profile menu → API Keys)<br>2. Leave all 10 capability checkboxes ticked (the default) — see the scope note below<br>3. Generate API Key and save it somewhere safe<br>4. Copy it |
MCP connections (connect in Claude settings):
| # | MCP | What it's for | How to connect |
|---|---|---|---|
| 1 | FiveAgents | Gateway for Gemini, email, and logging calls (Zernio + DataforSEO are now their own connectors — rows 2 & 7) | Required — configured in Step 7a |
| 2 | Zernio (required for publishing) | Social media publishing (Facebook, Instagram, LinkedIn, TikTok, X, …) — formerly Late, now its own MCP (no longer gateway-routed) | Custom Connector — URL https://mcp.zernio.com/mcp → sign in with your Zernio account (OAuth). Configured in Step 7a-ii |
| 3 | Notion | Content calendar management | Settings → Connected Apps → Notion → Authorize |
| 4 | Slack | Notifications | Settings → Connected Apps → Slack → Authorize |
| 5 | Gmail | Reading emails | Settings → Connected Apps → Gmail → Authorize |
| 6 | Google Calendar | Scheduling | Settings → Connected Apps → Google Calendar → Authorize |
| 7 | DataforSEO (optional — keyword research) | Keyword search volume + suggestions — used by research-strategy, trend-radar. Now its own MCP (no longer gateway-routed). Skip → those skills fall back to non-keyword research |
Custom Connector — URL https://mcp.dataforseo.com/mcp → authenticate with your DataforSEO API login + password (Basic Auth). Configured in Step 7a-iii |
| 8 | Windsor.ai (required) | Google Ads + GA4 + Meta Ads (Facebook + Instagram) analytics data — the universal source for all paid-ads + analytics reporting | 1. Sign up for a free account at https://windsor.ai/register<br>2. In Windsor dashboard, connect all three: Google Ads, GA4, and Meta Ads (Facebook Ads)<br>3. In Claude, go to Settings → Connected Apps → Windsor.ai → Authorize |
| 9 | Meta Ads (optional enhancement — limited rollout) | Meta's official MCP for direct Marketing API access. When available, skills prefer it for Meta data; otherwise the Windsor.ai connection above already covers Meta Ads. | Custom Connector — URL https://mcp.facebook.com/ads (configured in Step 7c). Skip without prejudice if your account doesn't have access yet — Windsor.ai already covers Meta Ads. |
| 10 | Canva | Campaign presentations and pitch decks | Settings → Connected Apps → Canva → Authorize |
| 11 | Apollo.io (business-ops) | Lead enrichment, contact search, sequence injection — used by apollo-lead-prospector, outreach-sequencer |
Settings → Connected Apps → Apollo.io → Authorize |
| 12 | Calendly (business-ops) | Kickoff scheduling and booking links — used by customer-onboarder, outreach-sequencer |
Settings → Connected Apps → Calendly → Authorize |
| 13 | Stripe (business-ops) | Invoice status and payment data — used by invoice-collector, financial-reporter |
Settings → Connected Apps → Stripe → Authorize (requires OAuth complete_authentication to unlock real tools) |
| 14 | Xero (business-ops) | Invoice sync + P&L pull — used by invoice-collector, financial-reporter |
Settings → Connected Apps → Xero → Authorize |
| 15 | PostHog (business-ops, optional) | Product-usage signals for churn scoring — used by churn-predictor. Skip → falls back to support-ticket + login-frequency only |
Settings → Connected Apps → PostHog → Authorize |
| 16 | Gamma (business-ops, optional) | Investor decks — used by investor-update-writer. Skip → updates render as plain markdown / email |
Settings → Connected Apps → Gamma → Authorize |
| 17 | n8n Cloud (business-ops, optional) | Builds the proof-of-concept automation that backs an inbound gig bid — used by n8n-workflow-builder. Skip → that skill is unconfigurable for this brand; the rest of the gig pipeline still runs |
Settings → Connected Apps → n8n Cloud → Authorize |
Present this overview to the user, then ask:
Ready to get started? We'll go through each step together.
Do not proceed to Step 3 until the user confirms they are ready.
Step 3 — Brand Name & Folder
Now we pick the slug that identifies this brand across the file system, env vars, and Notion DB names. The slug becomes part of paths like brands/{slug}/brand.md and env vars like ${SLUG_UPPER}_NOTION_DB, so keep it short and lowercase. Once chosen, the slug is permanent for this brand — changing it later means renaming a lot of files and re-pointing env vars.
Ask the user:
What is your brand name? This will become the folder name for all your brand assets (e.g. "acme" →
brands/acme/).
Important: Do NOT suggest existing brand names or show a list of previous brands. Simply ask for the brand name as a free-text input and wait for the user's answer.
Create the directory structure:
brands/{brand}/
├── brand.md
├── product.md
├── audience.md
├── competitors.md
├── funnel.md
├── sales.md ← Step 5g (v2.4.0 — required for apollo-lead-prospector / outreach-sequencer / proposal-generator)
├── customer-success.md ← Step 5h (v2.4.0 — required for customer-onboarder / churn-predictor)
├── finance.md ← Step 5i (v2.4.0 — required for invoice-collector / financial-reporter)
├── investors.md ← Step 5j (v2.4.0 — required for investor-update-writer; SKIP if no outside funding raised)
├── operations.md ← Step 5k (v2.4.0 — required for meeting-analyzer; SKIP if you don't process meeting transcripts)
├── logo.png ← Step 6
├── backgrounds/
└── design-system/ ← installed in Step 4b (Claude Design — OPTIONAL, recommended; stored locally, optionally also uploaded to fb.ai)
Social templates (Step 4c) and the media library (Step 4d) are not stored locally — they live on fb.ai (discovered at runtime via
fivebucks_list_templatesandfivebucks_list_media_folders). The design system (Step 4b) is stored locally indesign-system/and can optionally also be uploaded to the fb.ai brand-kit dashboard (then discoverable viafivebucks_get_brand_kit).
Also create:
outputs/{brand}/
outputs/{brand}/strategy/
Do not proceed to Step 4 until the brand name is confirmed and all directories are created.
Step 4 — Website Analysis
This is where the heavy lifting starts. The agent uses Playwright MCP to navigate your website and extract everything we can automatically — tagline, voice samples, brand colors, fonts, value propositions, CTAs, and even a first-pass at your target personas. The output of this step seeds brand.md and audience.md, which then feed every downstream skill (research-strategy, content-creation, creative-designer, etc.).
Expect ~5–10 minutes of mostly-watching while the agent crawls the site, plus 1–2 minutes of you reviewing the extracted personas and confirming they look right. The agent will ask one question (your website URL) and then drive the rest until it's ready for review.
Ask the user:
What is your website URL? (e.g. https://acme.com)
Use Playwright MCP to navigate to the site and extract content:
browser_navigateto the homepage — extract tagline, value propositions, hero copy, CTAsbrowser_snapshotto read the DOM and discover all top-level navbar links- Visit each navbar page (e.g. Pricing, Services, Portfolio, About, Contact, Blog) using
browser_navigate+browser_snapshotand extract key content from each
While browsing, also extract the brand's locale and visual identity:
- Locale — infer from the website's TLD (e.g.
.co.id→ Indonesia), language, physical address, phone number country code, or currency symbols on the page. Determine:- Currency code and symbol (e.g.
Rpfor IDR,SGD,$for USD) - Timezone (e.g.
Asia/Jakarta,Asia/Singapore) - Meta Ads USD exchange rate (Facebook spend is always USD — this rate converts to local currency) If ambiguous, confirm with the user.
- Currency code and symbol (e.g.
- Colors — use
browser_evaluateto inspect computed CSS styles on buttons, headings, and nav elements to find primary, secondary, and accent HEX codes:{ button: getComputedStyle(document.querySelector('button, .btn, [class*="btn"]') || document.body).backgroundColor, heading: getComputedStyle(document.querySelector('h1, h2') || document.body).color, nav: getComputedStyle(document.querySelector('nav, header') || document.body).backgroundColor, link: getComputedStyle(document.querySelector('a') || document.body).color } - Fonts — use
browser_evaluateto check<link>tags for Google Fonts URLs andfont-familyon headings/body:{ googleFonts: [...document.querySelectorAll('link[href*="fonts.googleapis.com"]')].map(l => l.href), headingFont: getComputedStyle(document.querySelector('h1, h2') || document.body).fontFamily, bodyFont: getComputedStyle(document.body).fontFamily }
If Playwright MCP is unavailable, blocked (e.g. Cloudflare protection), or fails to load the site, ask the user to paste the key content directly into the chat and provide their brand colors (HEX codes) and Google Font name manually.
Present the discovered colors and fonts to the user for confirmation:
I found these brand colors on your site: Primary:
#1A73E8, Secondary:#34A853, Accent:#FBBC04. Are these correct? Your site uses the Google Font Inter for headings and Roboto for body text. Should I use these?
If colors or fonts couldn't be extracted, ask the user directly:
- Brand Colors — "What are your brand colors? Please provide HEX codes (e.g. #1A73E8)."
- Google Font — "What Google Font does your brand use? Browse at https://fonts.google.com"
- Voice & Tone — if unclear from the website copy, ask the user to describe it
Using the analyzed data + any corrections from the user, generate two files:
brands/{brand}/brand.md
# {Brand Name}
## Tagline
{extracted or user-provided}
## Voice & Tone
- {infer from website copy — e.g. "Professional but approachable", "Bold and direct"}
## Colors (HEX codes)
- Primary: {extract HEX from website if possible, otherwise ask user for HEX code e.g. #1A73E8}
- Secondary: {extract HEX or ask user for HEX code}
- Accent: {extract HEX or ask user for HEX code}
## Locale
- Currency: {e.g. Rp, SGD, USD — infer from website/country}
- Timezone: {e.g. Asia/Jakarta, Asia/Singapore — infer from location}
- Meta USD exchange rate: {e.g. 16200 for IDR, 1.36 for SGD — USD to local currency}
## Approved Phrases
- {key phrases from the website}
## Do NOT Say
- {common mistakes to avoid — infer from brand positioning}
## Content Strategy
{written by brand-setup Step 4a — see below}
brands/{brand}/audience.md
# {Brand Name} — Target Audience
## Personas
### {persona-slug-1}
- **Title:** {e.g. Marketing Manager}
- **Pain Points:** {inferred from website messaging}
- **Goals:** {what they want to achieve}
- **Objections:** {why they might not buy}
### {persona-slug-2}
...
Generate 3-6 personas based on the website's messaging and target market.
Show the user each draft and let them review/edit before saving.
Do not proceed to Step 4b until the user has reviewed and confirmed brand.md and audience.md.
Step 4a — Content Strategy
Before setting up visual assets, capture how this brand will publish content. This single field drives social-calendar's planning mode at runtime — it must be set correctly now.
Ask the user:
What is your primary content channel?
- YouTube-first — you record YouTube videos weekly, then distribute clips to other platforms
- Static — you publish standalone posts, carousels, and stories directly to each platform
If YouTube-first, also ask:
Which platforms will you distribute clips to? (select all that apply: LinkedIn, Instagram, TikTok, Twitter/X, Facebook) Which of those are currently connected in Zernio? (we'll set account IDs in Step 7) How many clips per video per platform? (default: 1–2)
If Static, no additional questions — the existing 14-post weekly slot flow applies.
Write the result to brands/{brand}/brand.md under ## Content Strategy:
YouTube-first:
## Content Strategy
- **Primary channel:** youtube
- **Distribution platforms:** {user-selected list, e.g. LinkedIn, Instagram, TikTok (pending), Facebook (pending)}
- **Connected platforms:** {only platforms already connected, e.g. LinkedIn, Instagram}
- **Clips per video:** {user-provided or default 1–2} per platform
- **Cadence:** 1 YouTube video/week → clips distributed to connected platforms
Static:
## Content Strategy
- **Primary channel:** static
- **Distribution platforms:** {connected platforms from Step 7, e.g. LinkedIn, Instagram, Facebook}
- **Cadence:** 14 posts/week (Mon–Sat fixed slots)
Important rules:
- Never hardcode platform names — always write exactly what the user confirmed
social-calendarreadsPrimary channelat runtime to switch between YouTube-First Mode and Static Mode- When a new platform is connected later, the user updates
Connected platformsinbrand.md— no skill changes required Distribution platforms= all platforms the brand intends to use (including not-yet-connected);Connected platforms= subset currently live
Do not proceed to Step 4b until ## Content Strategy is written to brand.md and confirmed by the user.
Step 4b — Claude Design System (OPTIONAL — recommended)
Optional but recommended. A Claude Design system is a structured visual identity (colors, fonts, components, spacing) that lives at brands/{brand}/design-system/. When present, every visual output across the plugin (social images, decks, mockups, email templates, ad creatives) reads from it as the authoritative source — guaranteeing consistency. When absent, all those outputs fall back to the colors and fonts captured in brand.md from Step 4, which works fine but tends to drift visually over time.
Expect ~10 minutes if you decide to install one (you'll create it at claude.ai/design first, then copy it into the brand folder; optionally you can also upload it to the fb.ai brand-kit dashboard). Skip it if you're early-stage and brand consistency isn't a top priority — you can always re-run brand-setup later to add it.
Anthropic ships Claude Design — a visual brand-system builder at https://claude.ai/design. When installed, it becomes the source of truth for visual identity (fonts, colors, components, spacing) and is preferred by every downstream skill that produces visuals. It is recommended for the strongest brand consistency, but it is not required — Step 9c will detect whether the design system was installed and skills will fall back to the colors/fonts/voice in brands/{brand}/brand.md (plus Gemini + Pillow rendering) when it isn't there.
fb.ai context — show this before any fivebucks.ai link. fivebucks.ai (fb.ai) is a paid product. Never hand the user a fivebucks.ai URL or ask them to upload anything there without first explaining what it is — Steps 4b-D, 4c, and 4d all link to fb.ai, so the user should already have this context by the time they see the first link:
About fb.ai (fivebucks.ai): it's an optional paid subscription that stores your design system and social templates in the cloud, so Claude can automate your branded social posting — generating and publishing fully on-brand Carousels, Stories, and IG / FB / LinkedIn posts for you. Your design system also lives locally for free (this step), so you can start without paying; the fb.ai plan adds the hosted social templates and lets posts render in your exact brand colors and fonts. You can skip fb.ai entirely — Claude falls back to the colors and fonts in
brand.mdplus Gemini rendering, which stays fully supported.
Ask the user:
Want me to walk you through creating a Claude Design system for your brand? It's the strongest way to keep every image, post, and deck visually consistent. We can also skip and rely on the brand colors / fonts you already confirmed in
brand.md— skills will fall back gracefully.
If the user says skip, acknowledge and move on to Step 4c. If yes, walk them through the installation:
Step A — Create the design system:
- Open https://claude.ai/design in your browser
- Go to Design System and click Create
- On the "Set up your design system" page, fill in the attach form:
- Company name and blurb — your brand name + a one-line description of what you do
- Provide examples (all optional) — attach any design resources you already have: link a GitHub repo, link code from your computer, upload a .fig (Figma) file, and/or add fonts, logos and assets (drop your logo file here)
- Any other notes? — paste your brand colors (HEX codes from
brands/{brand}/brand.mdwe just generated), your heading + body fonts (the Google Fonts from Step 4), and your brand voice / aesthetic notes frombrand.md- Click Continue to generation (top-right)
- On the generation prompt, paste the prompt below — it points Claude at your live site so it can read it and validate / fill in the visual details automatically. Then iterate with Claude until the design system feels on-brand.
Compose the generation prompt before showing the user — substitute every {...} from brands/{brand}/brand.md plus the website URL the user gave in Step 4, so the user sees only finished text to copy-paste:
| Placeholder | Source |
|---|---|
{WEBSITE_URL} |
the website URL the user gave in Step 4 |
{BRAND_NAME} |
proper-cased brand name (from Step 3 — not the slug) |
{ONE_LINER} |
one-line description of the brand (from brand.md Tagline / website) |
{PRIMARY_HEX}, {SECONDARY_HEX}, {ACCENT_HEX} |
brand.md Colors section |
{HEADING_FONT}, {BODY_FONT} |
Google Fonts identified in Step 4 |
{VOICE_TONE} |
brand.md Voice & Tone section, summarized to one short phrase |
Then present it like this:
Here's a prompt to paste into the Claude Design generation box — it points Claude at your live site so it can validate and fill in the details for you automatically:
This is my website: {WEBSITE_URL} — please read it and use it to validate and fill in my visual identity: color palette, typography, components, spacing, and overall aesthetic. Here's what I already know — confirm it against the site and refine wherever the site disagrees: - Brand: {BRAND_NAME} — {ONE_LINER} - Colors: primary {PRIMARY_HEX}, secondary {SECONDARY_HEX}, accent {ACCENT_HEX} - Typography: {HEADING_FONT} for headings, {BODY_FONT} for body - Voice / aesthetic: {VOICE_TONE} Build a complete, reusable design system: color tokens, a type scale, buttons, cards, headers / nav, and spacing — all matching the site.Step B — Export the design system:
- Click Share → Download Project as .zip
- Unzip the file on your computer (you'll get a folder named something like "{Brand Name} Design System")
Step C — Tell me the path to the unzipped folder; I'll request access and copy it into place:
- Leave the unzipped folder wherever it is — e.g.
~/Downloads/Acme Design System. You do not need to move it into the project.- Tell me the path. I'll request read access to that exact folder — you'll get a one-time approval prompt; click Allow. Then I copy its contents into
brands/{brand}/design-system/automatically (no renaming required).
Then ask:
What's the path to your unzipped Claude Design folder? (e.g.
~/Downloads/Acme Design System, or a project-relative path like./design-temp/Acme Design Systemif you already dropped it inside the project.)
Request access to the user-provided path before reading it. The path the user just gave is the one to mount — request access to that exact folder, and the user approves it in a one-time prompt:
mcp__cowork__request_cowork_directory(path="<the path the user gave>")
This is required in the Cowork harness, where filesystem access is sandboxed to approved directories. Request the folder the user named — not a guessed default — so the approval prompt matches what they typed. (In local Claude Code the filesystem is already accessible and this tool is absent; skip the request and copy directly.)
If the user declines the access prompt, ask them to move the unzipped folder inside the project folder (anywhere within the mount), then copy from there.
Once access is approved, copy the folder using Python:
import shutil
from pathlib import Path
raw = user_input.strip().strip('"').strip("'")
src = Path(raw).expanduser().resolve()
assert src.exists(), f"Source not found: {src}"
assert src.is_dir(), f"Not a directory: {src}"
assert any(src.iterdir()), f"Source folder is empty: {src}"
# Handle the common nested-zip case — sometimes unzipping creates a wrapper
# folder containing one inner folder with the same name. Reach inside it.
contents = list(src.iterdir())
if len(contents) == 1 and contents[0].is_dir():
src = contents[0]
dst = Path("brands") / brand / "design-system"
if dst.exists():
shutil.rmtree(dst) # idempotent re-run support
shutil.copytree(src, dst)
Confirm to the user:
✅ Copied to
brands/{brand}/design-system/. Your original folder is untouched and can be deleted whenever you like.
If the user-provided path is invalid (not found, not a directory, or empty), tell them what was wrong and ask again — don't fall back to "manually rename" since the whole point of this step is to remove that burden.
Verification (after the copy succeeds):
Read brands/{brand}/design-system/index.html (or whatever entry file is present — index.html, index.tsx, App.tsx, README.md, etc.) to confirm the design system contains the brand's colors and typography. If nothing identifiable is found, surface the file listing to the user and ask whether the export looks right — Claude Design's export structure can vary.
If colors/fonts in the Claude Design system differ from brands/{brand}/brand.md (e.g. user refined them in Claude Design), update brand.md to match. When the design system is installed, it is authoritative; when it isn't, brand.md remains the source of truth on its own.
Step D — (OPTIONAL) Also upload to the fb.ai brand-kit dashboard.
After the local copy succeeds, ask the user whether they also want to upload the design system to fb.ai. This is optional — the local design-system/ folder is fully functional on its own. Uploading to fb.ai lets fb.ai-rendered social templates (Step 4c) pick up the brand's colors/fonts natively.
Want to also upload this design system to fb.ai? It's optional — your local copy already works everywhere. Uploading lets the fb.ai social templates render with your brand colors/fonts automatically. (Needs a paid fb.ai plan.)
If the user says skip, that's fine — move on to Step 4c. If yes, walk them through it:
- Go to https://www.fivebucks.ai/dashboard/api-keys and sign in — generate your API Key and save it somewhere safe. (Skip if you have already generated your API Key.)
- Go to https://www.fivebucks.ai/dashboard/social-posts/brand-kit — click Upload Design System Zip and upload the ZIP file.
Proceed to Step 4c whether or not the design system was installed — Step 9c will record the actual installed/missing state in CLAUDE.md, and skills will branch accordingly at runtime.
Step 4c — Social Templates (OPTIONAL — recommended)
Optional but recommended for brands publishing regularly to IG / FB / LinkedIn. These are Claude Design templates that render polished, fully-branded posts — far better than the Gemini + Pillow fallback. They live on fb.ai (you upload them in the fb.ai dashboard), and content-generator / creative-designer render them at runtime via the gateway fivebucks_* tools. If you skip any (or all), nothing breaks — those channels fall back to the Gemini + Pillow image-path, which stays fully supported.
Prerequisite: a paid fb.ai subscription and the brand's fb.ai API key (FIVEBUCKS_API_KEY — set up in Step 7). Without it, skip Step 4c entirely and rely on the fallback.
Four template types can be installed (all optional — set up only the channels you publish on):
Type (fivebucks type) |
Channel | Canvas | Structure |
|---|---|---|---|
meta-carousel |
IG / FB carousel | 4:5 | 6 slides: Cover + 4 value slides + CTA |
meta-story |
IG / FB Stories | 9:16 | 6 slides (Hook→Problem→Solution→Proof→Offer→CTA) × 3 directions A/B/C |
linkedin-post |
LinkedIn single-image feed | 4:5 | 1 slide, 3 directions A/B/C (Hook Headline / Stat Hero / Pull Quote) |
meta-post |
IG / FB single-image feed | 4:5 | 1 slide, 3 directions A/B/C (Hero Visual / Quote Card / Listicle Teaser) |
Expect ~5 min per template (author in Claude Design → export → upload to the fb.ai dashboard).
Ask the user:
Want to set up any of the optional social templates? They give your Carousel / Story / LinkedIn / IG-FB posts a fully-branded layout instead of a generic Gemini background. They need a paid fb.ai plan. We can skip any channel — those just use the Gemini + Pillow fallback.
If the user skips all, move on to Step 5. Otherwise run only the sub-steps for the channels they want.
Install flow (identical for all four templates)
Step A — Author in Claude Design. Open https://claude.ai/design, create a project, and paste the type-specific prompt from the matching sub-step below (4c-i … 4c-iv). Iterate until it looks on-brand. The prompt already encodes the native-image contract (below) — keep that part intact.
Step B — Export. In Claude Design: Share → Download Project as .zip.
Step C — Upload to the fb.ai dashboard.
- Go to https://www.fivebucks.ai/dashboard/api-keys and sign in — generate your API Key and save it somewhere safe. (Skip if you have already generated your API Key.)
- Go to https://www.fivebucks.ai/dashboard/social-posts/templates — click Upload Templates and upload the ZIP file.
fb.ai unpacks the ZIP, reads the EDITMODE block into a manifest (editable fields + image slots + slides), and detects the template type from its slug. (Template upload is dashboard-only — there is no gateway upload tool.)
Step D — Verify via the gateway.
Use gateway MCP tool fivebucks_list_templates:
- fiveagents_api_key: ${FIVEAGENTS_API_KEY}
Confirm an entry with the expected type appears, then call fivebucks_get_template (template_id) and confirm manifest.fields include the copy keys + image slots for that type. If absent, the upload didn't land — re-check the dashboard upload. No local copy is kept — the template lives only on fb.ai; detection everywhere uses fivebucks_list_templates.
Native-image contract (already baked into every prompt below — keep it intact)
Every template renders its own photos and tint overlay natively from its EDITMODE image slots — fb.ai does not composite images server-side. Each photo slot is three EDITMODE keys:
"<slot>_image": "", // "" or "media:{fileId}" — fb.ai resolves it at render time
"<slot>_image_position": "center", // center | left | right | top | bottom
"<slot>_image_fit": "cover" // cover | contain
and the template JSX renders, behind the content:
{tweaks.<slot>_image && (<img src={tweaks.<slot>_image} alt="" style={{position:'absolute',inset:0,width:'100%',height:'100%',objectFit:tweaks.<slot>_image_fit||'cover',objectPosition:tweaks.<slot>_image_position||'center',zIndex:0}} />)}
{tweaks.<slot>_image && (<div style={{position:'absolute',inset:0,background: THEME==='light' ? 'rgba(255,255,255,0.5)' : 'rgba(0,0,0,0.4)',zIndex:1}} />)}
with the slide's text/content wrapper at position:relative; zIndex:2 so copy always sits above the photo + overlay. Rules:
- Logo is bundled at
assets/logo.pngand referenced directly (<img src="assets/logo.png">). It must NOT appear in the EDITMODE JSON under any key. - No hardcoded photo paths in EDITMODE, no base64 data URIs, no bundled demo photos. Empty
_imageslots render the template's own gradient/placeholder. - Overlay opacity is fixed: dark =
rgba(0,0,0,0.4), light =rgba(255,255,255,0.5). The theme indicator (_template_theme, or_theme_A/B/Cfor meta-story) sets the template's ownTHEMEconstant for the overlay.
Export-marker contract (already baked into every prompt below — keep it intact)
fb.ai's server-side renderer screenshots one PNG per exportable slide by locating elements by data-export-id. It captures each marked element at the manifest's canvas size (e.g. 1080×1350 / 1080×1920), pinning it to the viewport and hiding everything else first — so the marked element must BE the full-size artboard, and the renderer no longer cares whether the preview parks or scales it. Follow these rules or the wrong elements get exported:
- Mark the full-size (1:1) artboard, exactly once. Put
data-export-idon the outermost element that is, in its own coordinate space, the complete slide at canvas dimensions (its own background + content). Do not also mark a second copy (e.g. a scaled design-canvas preview AND a hidden export frame) — duplicate markers make the renderer register and screenshot the wrong count. Pick one element per slide. - Static string literals only. Write the id directly, e.g.
data-export-id="A-0". Never a JSX expression (data-export-id={A-${i}}) and never a template literal trapped in quotes (data-export-id="${dir}-${i}"). The uploader scans the HTML statically at upload time; both forms are invisible to it (or register a junk literal) and corrupt the slide manifest. - Renderable, never
display:none. The marked element may be visually hidden for the preview via off-screen offset,transform: scale(),opacity, orvisibility:hidden— the renderer un-scales and relocates it — butdisplay:noneremoves it from layout and yields a blank capture. For multi-direction single-slide templates, gate the inactive directions however you like, but ensure the direction that will be exported is notdisplay:none. - All artboards present in the DOM regardless of the active
_direction/direction; the renderer reads that key from EDITMODE to choose which subset to screenshot.
The four type-specific prompts follow. Each ends with the EDITMODE block; the image-slot trio + native render rules above apply to all of them.
4c-i. Meta Carousel Template (Instagram + Facebook, 4:5)
Step A — Generate the template in Claude Design:
The agent gives the user a fully-composed, copy-pasteable prompt to drop into Claude Design. Claude Design produces a ready-to-use React + Babel template app with an EDITMODE block — the user iterates on the design, not the code.
Compose the prompt before showing the user. Substitute the placeholders below from brands/{brand}/brand.md. The user must see only the finished prompt — no {...} markers left over.
| Placeholder | Source |
|---|---|
{BRAND_NAME} |
proper-cased brand name (from Step 3 — not the slug) |
{PRIMARY_HEX}, {SECONDARY_HEX}, {ACCENT_HEX} |
brand.md Colors section |
{HEADING_FONT}, {BODY_FONT} |
Google Fonts identified in Step 4 |
{VOICE_TONE} |
brand.md Voice & Tone section, summarized to one short phrase |
Then present to the user, framed exactly like this:
Here's a prompt I've put together for your Meta Carousel template. Copy and paste it into Claude Design:
- Open https://claude.ai/design in your browser
- Create a new project (name it whatever you like — e.g. "Acme Carousel Template")
- Paste the prompt below into Claude Design's chat:
"I need a configurable Instagram + Facebook carousel template for {BRAND_NAME}. 4:5 portrait, 1080×1350, 6 slides total: 1 Cover slide + 4 sign / value slides + 1 CTA slide. Build it as a React + Babel app rendered from a single index.html (use Babel standalone via CDN — no build step). The agent will modify copy programmatically per post, so structure matters more than specific sample copy. The brand uses {PRIMARY_HEX} as primary, {SECONDARY_HEX} as secondary, {ACCENT_HEX} as accent. Headings are set in {HEADING_FONT}, body in {BODY_FONT}. The aesthetic is {VOICE_TONE}. Include the {BRAND_NAME} brand logo subtly in a corner of every slide. Each slide should also have a small page indicator (e.g. '1/6', '2/6'), and the 4 sign slides should have a large kicker numeral on the left ('01', '02', '03', '04'). CRITICAL CONTRACT — wrap all editable copy in a single JSON object inside index.html, marked with /*EDITMODE-BEGIN*/ and /*EDITMODE-END*/ comment markers, exactly like this:window.__TWEAK_DEFAULTS = /EDITMODE-BEGIN/{ "accent": "purple", "footer": true, "coverVariant": "type", "bodyVariant": "allnumbers", "slideTheme_1": "dark", "slideTheme_2": "cream", "slideTheme_3": "cream", "slideTheme_4": "cream", "slideTheme_5": "cream", "slideTheme_6": "dark", "_template_theme": "dark", "handle": "@{BRAND_NAME_SLUG}", "hashtag": "#YourTag", "cover_eyebrow": "sample eyebrow", "cover_title": "sample cover headline", "cover_sub": "sample cover subline", "s2_kicker": "01", "s2_title": "...", "s2_body": "...", "s2_image": "", "s2_image_position": "center", "s2_image_fit": "cover", "s3_kicker": "02", "s3_title": "...", "s3_body": "...", "s3_image": "", "s3_image_position": "center", "s3_image_fit": "cover", "s4_kicker": "03", "s4_title": "...", "s4_body": "...", "s4_image": "", "s4_image_position": "center", "s4_image_fit": "cover", "s5_kicker": "04", "s5_title": "...", "s5_body": "...", "s5_image": "", "s5_image_position": "center", "s5_image_fit": "cover", "cta_eyebrow": "...", "cta_title": "...", "cta_sub": "...", "cta_button": "..." }/EDITMODE-END/;
Use these EXACT key names — the agent substitutes copy by parsing the JSON between the markers and writing back. Optional supporting keys per sign slide are welcome (e.g. s2_pullquote, s3_stat_value/s3_stat_label, s5_before/s5_after) — the agent can populate them when the post copy provides matching fields, otherwise leave defaults. THEME INDICATOR — set `_template_theme` to `\"dark\"` or `\"light\"` to match the value slides' (s2–s5) background scheme. The template uses it to pick the native overlay color (dark = rgba(0,0,0,0.4), light = rgba(255,255,255,0.5)). fb.ai surfaces it as `manifest.theme` but composites nothing — the template owns the overlay. EXPORT MARKERS — fb.ai's renderer screenshots one PNG per slide by `data-export-id`, capturing each marked element at full 1080×1350. Put the attribute on the **full-size (1:1) artboard element** — the 1080×1350 slide itself, NOT the scaled-down preview wrapper your design-canvas rail renders it in (the renderer un-scales whatever it's told to capture, but mark the real artboard so it captures the whole slide). Use a **static string literal**, marked exactly once per slide: `data-export-id=\"A-0\"`–`\"A-5\"` for direction A, equivalently for B and C (18 artboards total). Never a JSX expression (`data-export-id={\`A-${i}\`}`) or a quoted template literal (`data-export-id=\"${dir}-${i}\"`) — the uploader scans statically and both forms corrupt the manifest. All 18 artboards stay in the DOM (never `display:none` the one being exported); the renderer reads `_direction` from EDITMODE to pick which 6. IMAGE SLOTS — the 4 value slides (s2–s5) each take a user-swappable photo via `s{n}_image` / `s{n}_image_position` / `s{n}_image_fit` (defaults empty / center / cover). Render them NATIVELY per the native-image contract above: when `tweaks.s{n}_image` is set, draw a full-bleed `<img>` at zIndex:0 (objectFit/objectPosition from the companion keys), a theme overlay div at zIndex:1, and keep the slide's content wrapper at position:relative; zIndex:2; when empty, show the slide's own gradient/placeholder. Cover (s1) and CTA (s6) are photo-free. The brand logo is bundled at `assets/logo.png` and referenced directly — never in EDITMODE. No base64 data URIs, no bundled photo/demo assets, no unused fonts. Use the brand colors and fonts above. The template's sample copy will be replaced at runtime — don't worry about it being final."Iterate with Claude inside Claude Design until the 6 slide layouts look on-brand — you can ask for tweaks like "make the cover darker" or "swap the gradient direction on slide 4". Don't worry about the sample copy; it gets replaced per post at runtime. Once the layout looks good, let me know and we'll move to export.
Wait for the user to confirm they're happy with the template before exporting. Then run the shared install flow above (Step B export → Step C upload to the fb.ai dashboard → Step D verify via fivebucks_list_templates). The meta-carousel manifest should expose cover_*, s2_*–s5_* (including each _image trio), and cta_* fields. If the user skips this template, that's fine — IG/FB carousels fall back to the Gemini + Pillow image-path.
4c-ii. Meta Story Template (Instagram + Facebook, 9:16)
Step A — Generate the template in Claude Design:
Same pattern as Step 4c-i: the agent composes a copy-pasteable prompt and gives it to the user. Claude Design produces a React + Babel template app with an EDITMODE block; the user iterates on the design.
Compose the prompt before showing the user. Substitute the same placeholders as in 4c-i (from brands/{brand}/brand.md): {BRAND_NAME}, {PRIMARY_HEX}, {SECONDARY_HEX}, {ACCENT_HEX}, {HEADING_FONT}, {BODY_FONT}, {VOICE_TONE}. The user must see only the finished prompt — no {...} markers left over.
Then present to the user, framed exactly like this:
Here's a prompt I've put together for your Meta Story template. Copy and paste it into Claude Design:
- Open https://claude.ai/design in your browser
- Create a new project (name it whatever you like — e.g. "Acme Story Template")
- Paste the prompt below into Claude Design's chat:
"I need a configurable Instagram + Facebook Story template for {BRAND_NAME}. 9:16 vertical, 1080×1920, 6 slides total following a Hook → Problem → Solution → Proof → Offer → CTA narrative. Build it as a React + Babel app rendered from a single entry HTML (Babel standalone via CDN — no build step). Three direction styles (A / B / C) sharing the same copy but different visual treatments. The brand uses {PRIMARY_HEX} as primary, {SECONDARY_HEX} as secondary, {ACCENT_HEX} as accent. Headings are set in {HEADING_FONT}, body in {BODY_FONT}. Aesthetic: {VOICE_TONE}. Include the {BRAND_NAME} brand logo on every slide. Respect IG safe zones (top ~14% for username/header, bottom ~14% for reply bar / Reel UI). Direction styles: - **Direction A — Spotlight Dark:** brand-led campaigns. Eyebrow → headline → divider pill, dark backgrounds, accent color leading. - **Direction B — Editorial Stat:** when a single big number is the hero. Oversized stat as the centerpiece. - **Direction C — Cream Press:** light, magazine-style. For case studies, testimonials, founder posts. CRITICAL CONTRACT — wrap all editable copy in a JSON object inside the entry HTML, marked with /*EDITMODE-BEGIN*/ and /*EDITMODE-END*/ comment markers, using these EXACT key names per slide:/EDITMODE-BEGIN/{ "_direction": "A", "_theme_A": "dark", "_theme_B": "dark", "_theme_C": "light", "handle": "@{BRAND_NAME_SLUG}", "s1_eyebrow": "...", "s1_headline_pre": "...", "s1_headline_accent": "...", "s1_sub": "...", "s1_live": "...", "s1_big": "...", "s1_big_unit": "...", "s2_eyebrow": "...", "s2_headline": "...", "s2_pain1": "...", "s2_pain2": "...", "s2_pain3": "...", "s2_image": "", "s2_image_position": "center", "s2_image_fit": "cover", "s3_eyebrow": "...", "s3_headline_pre": "...", "s3_headline_accent": "...", "s3_sub": "...", "s3_image": "", "s3_image_position": "center", "s3_image_fit": "cover", "s4_eyebrow": "...", "s4_headline": "...", "s4_stat1_num": "...", "s4_stat1_lbl": "...", "s4_stat2_num": "...", "s4_stat2_lbl": "...", "s4_stat3_num": "...", "s4_stat3_lbl": "...", "s4_stat4_num": "...", "s4_stat4_lbl": "...", "s4_quote": "...", "s4_quote_author": "...", "s4_image": "", "s4_image_position": "center", "s4_image_fit": "cover", "s5_eyebrow": "...", "s5_headline": "...", "s5_b1": "...", "s5_b2": "...", "s5_b3": "...", "s5_b4": "...", "s5_pill": "...", "s5_image": "", "s5_image_position": "center", "s5_image_fit": "cover", "s6_eyebrow": "...", "s6_headline_pre": "...", "s6_headline_accent": "...", "s6_sub": "...", "s6_cta": "...", "s6_url": "..." }/EDITMODE-END/;
THEME INDICATORS — set `_theme_A`, `_theme_B`, `_theme_C` to `\"dark\"` or `\"light\"` to match each direction's background scheme (defaults A=dark, B=dark, C=light). The template uses the active direction's theme to pick the native overlay color (dark = rgba(0,0,0,0.4), light = rgba(255,255,255,0.5)). fb.ai surfaces these as `manifest.theme` but composites nothing — the template owns the overlay. EXPORT MARKERS — fb.ai's server-side renderer screenshots one PNG per slide by `data-export-id`, capturing each marked element at full 1080×1920 (it pins the element to the viewport and hides everything else first). Follow these rules exactly: 1. **Enumerate all 18 full-size artboards, each marked exactly once.** Provide one `<div>` per slide that IS the complete 1080×1920 artboard (its own background + content), with a static `data-export-id`: `\"A-0\"`–`\"A-5\"` · `\"B-0\"`–`\"B-5\"` · `\"C-0\"`–`\"C-5\"` (3 directions × 6 slides = 18). All 18 stay in the DOM regardless of `_direction`; the renderer reads `_direction` from EDITMODE to pick which 6 to export. 2. **Mark the export artboard, not a second copy.** If your design-canvas shows scaled previews of the same slides, those preview copies MUST NOT carry `data-export-id` — only the 18 full-size artboards do. (A shared `StoryFrame` used for both preview and export would otherwise register 36 markers and corrupt the slide count.) 3. **Static string literals only** — write `data-export-id=\"A-0\"` directly on the wrapper `<div>`. Never a JSX expression (`data-export-id={\`A-${i}\`}`) and never a template literal trapped in quotes (`data-export-id=\"${dir}-${i}\"`). The uploader scans statically; both forms are invisible to it (or register a junk literal) and corrupt the manifest. 4. **Keep them renderable.** The 18 full-size artboards may sit in an off-screen / `opacity:0` container so they don't clutter the preview — the renderer relocates and captures them at full size regardless — but never `display:none` them (that yields a blank capture). IMAGE SLOTS — body slides s2–s5 (Problem → Proof) each take a user-swappable photo via `s{n}_image` / `s{n}_image_position` / `s{n}_image_fit` (defaults empty / center / cover), shared across directions A/B/C. Render them NATIVELY per the native-image contract above: full-bleed `<img>` at zIndex:0 (objectFit/objectPosition from the companion keys), a direction-themed overlay div at zIndex:1 (per `_theme_A`/`_theme_B`/`_theme_C`), content wrapper at position:relative; zIndex:2; when empty show the slide's own gradient/placeholder. s1 (Hook) and s6 (CTA) are photo-free. The brand logo is bundled at `assets/logo.png` and referenced directly — never in EDITMODE. No base64 data URIs, no bundled photo/demo assets, no unused fonts. Use brand colors and fonts. Sample copy will be replaced at runtime."Iterate with Claude inside Claude Design until the 6 slide layouts (Hook → Problem → Solution → Proof → Offer → CTA) and the 3 directions (A/B/C) all look on-brand. Don't worry about the sample copy; it gets replaced per post at runtime. Once the layout looks good, let me know and we'll move to Step B (export).
Wait for the user to confirm they're happy with the template before exporting. Then run the shared install flow above (Step B export → Step C upload to the fb.ai dashboard → Step D verify via fivebucks_list_templates). The meta-story manifest should expose _direction + _theme_A/B/C, s1_*–s6_* (with the _image trio on s2–s5), and the per-direction data-export-id slides. If the user skips this template, IG/FB Stories fall back to the Gemini + Pillow image-path.
4c-iii. LinkedIn Post Template (LinkedIn, 4:5)
Step A — Generate the template in Claude Design:
Same pattern as Step 4c-i and 4c-ii: the agent composes a copy-pasteable prompt and gives it to the user. Claude Design produces a React + Babel template app with an EDITMODE block; the user iterates on the design.
LinkedIn single-image feed posts are one slide, not a multi-slide deck. Default canvas is 1080×1350 (4:5 portrait) — current LinkedIn feed guidance favors this aspect for maximum mobile screen real estate. The template supports three directions (A/B/C) sharing one EDITMODE block but rendering different visual treatments per post.
Compose the prompt before showing the user. Substitute the same placeholders as in 4c-i / 4c-ii (from brands/{brand}/brand.md): {BRAND_NAME}, {PRIMARY_HEX}, {SECONDARY_HEX}, {ACCENT_HEX}, {HEADING_FONT}, {BODY_FONT}, {VOICE_TONE}, {BRAND_NAME_SLUG}. The user must see only the finished prompt — no {...} markers left over.
Then present to the user, framed exactly like this:
Here's a prompt I've put together for your LinkedIn Post template. Copy and paste it into Claude Design:
- Open https://claude.ai/design in your browser
- Create a new project (name it whatever you like — e.g. "Acme LinkedIn Post Template")
- Paste the prompt below into Claude Design's chat:
"I need a configurable LinkedIn single-image feed post template for {BRAND_NAME}. 4:5 portrait, 1080×1350, ONE slide. Build it as a React + Babel app rendered from a single index.html (use Babel standalone via CDN — no build step). The agent will modify copy programmatically per post, so structure matters more than specific sample copy. The brand uses {PRIMARY_HEX} as primary, {SECONDARY_HEX} as secondary, {ACCENT_HEX} as accent. Headings are set in {HEADING_FONT}, body in {BODY_FONT}. The aesthetic is {VOICE_TONE}. Include the {BRAND_NAME} brand logo subtly in a corner of the slide and the handle / URL as a small footer. Three direction styles (A / B / C) sharing the same EDITMODE keys but different visual treatments — the agent picks a direction per post via the `direction` key: - Direction A — Hook Headline (text-led): large oversized headline as the centerpiece, eyebrow chip above, 1–2 lines of supporting body underneath, accent-color CTA pill at bottom. Best for narrative / opinion posts. - Direction B — Stat Hero (numeric): an extremely large stat_value as the visual hero (think 60–70% of slide height), short stat_label underneath, eyebrow above, small body line at bottom. Best for proof / results posts. - Direction C — Pull Quote (editorial): large quote in serif/italic, attribution line beneath, optional small headshot circle, cream / light background. Best for testimonials, founder posts, press mentions. CRITICAL CONTRACT — wrap all editable copy in a single JSON object inside index.html, marked with /*EDITMODE-BEGIN*/ and /*EDITMODE-END*/ comment markers, using these EXACT key names:window.__TWEAK_DEFAULTS = /EDITMODE-BEGIN/{ "direction": "A", "accent": "purple", "_template_theme": "dark", "handle": "@{BRAND_NAME_SLUG}", "hashtag": "#YourTag", "eyebrow": "sample eyebrow", "headline": "sample LinkedIn headline that hooks the reader", "body": "sample supporting body line, ~1–2 sentences max.", "stat_value": "78%", "stat_label": "of teams say X", "quote": "sample quote that fits a pull-quote treatment.", "attribution": "Jane Doe, CEO", "bg_image": "", "bg_image_position": "center", "bg_image_fit": "cover", "headshot_image": "", "cta_text": "Learn more", "cta_button": "Read the full post" }/EDITMODE-END/;
Use these EXACT key names — the agent substitutes copy by parsing the JSON between the markers and writing back. The `direction` key (A / B / C) selects which visual treatment renders. Unused keys for a given direction (e.g. stat_value when direction='A', or quote when direction='B') should be ignored gracefully, not shown. THEME INDICATOR — set `_template_theme` to `\"dark\"` or `\"light\"` to match directions A & B (which receive user photos). The template uses it to pick the native overlay color (dark = rgba(0,0,0,0.4), light = rgba(255,255,255,0.5)). fb.ai surfaces it as `manifest.theme` but composites nothing — the template owns the overlay. EXPORT MARKERS — fb.ai's renderer screenshots the active direction by `data-export-id`, capturing it at full 1080×1350. Each direction artboard's outer `<div>` (the full-size 1:1 slide) carries a **static string literal** `data-export-id`, e.g. `data-export-id=\"A · Hook Headline\"`, `data-export-id=\"B · Stat Hero\"`, `data-export-id=\"C · Pull Quote\"` (name them to match the direction label). Never a JSX expression or a quoted template literal (`data-export-id=\"${dir}\"`) — the uploader scans statically. All three artboards stay in the DOM; gate the inactive directions with `visibility:hidden`, `opacity:0`, or an off-screen offset — **not** `display:none`, which yields a blank capture if that direction is the one exported. IMAGE SLOTS — directions A & B take a full-bleed user photo via `bg_image` / `bg_image_position` / `bg_image_fit`; direction C takes an optional circular `headshot_image`. Render them NATIVELY per the native-image contract above: full-bleed `<img>` at zIndex:0 (objectFit/objectPosition from the companion keys), theme overlay div at zIndex:1, content wrapper at position:relative; zIndex:2; when `bg_image` is empty show the gradient/placeholder. Direction C must look balanced when `headshot_image` is empty. The brand logo is bundled at `assets/logo.png` and referenced directly — never in EDITMODE. No base64 data URIs, no bundled photo/demo assets, no unused fonts. SAFE ZONES — LinkedIn doesn't crop feed images, but the right ~80px and bottom ~80px can collide with the like/comment overlay on hover preview. Keep critical text and logo at least 80px from the right edge and bottom edge. Use the brand colors and fonts above. The template's sample copy will be replaced at runtime — don't worry about it being final."Iterate with Claude inside Claude Design until all three directions (A/B/C) look on-brand. Don't worry about the sample copy; it gets replaced per post at runtime. Once the layout looks good, let me know and we'll move to Step B (export).
Wait for the user to confirm they're happy with the template before exporting. Then run the shared install flow above (Step B export → Step C upload to the fb.ai dashboard → Step D verify via fivebucks_list_templates). The linkedin-post manifest should expose direction, _template_theme, the copy keys (eyebrow/headline/body/stat_value/stat_label/quote/attribution/cta_text/cta_button), and the bg_image trio (+ optional headshot_image). If the user skips this template, LinkedIn posts fall back to the Gemini + Pillow image-path.
4c-iv. Meta Post Template (Instagram + Facebook, 4:5)
Step A — Generate the template in Claude Design:
Same pattern as Steps 4c-i through 4c-iii: the agent composes a copy-pasteable prompt and gives it to the user. Claude Design produces a React + Babel template app with an EDITMODE block; the user iterates on the design.
Instagram and Facebook single-image feed posts are one slide, not a multi-slide deck (carousels are covered by Step 4c-i). Default canvas is 1080×1350 (4:5 portrait) — current 2026 IG and FB feed guidance both converge on this aspect ratio for maximum mobile screen real estate (IG's profile grid also defaults to 4:5 thumbnails as of 2026). The template supports three directions (A/B/C) sharing one EDITMODE block but rendering different visual treatments per post.
Compose the prompt before showing the user. Substitute the same placeholders as in 4c-i / 4c-ii / 4c-iii (from brands/{brand}/brand.md): {BRAND_NAME}, {PRIMARY_HEX}, {SECONDARY_HEX}, {ACCENT_HEX}, {HEADING_FONT}, {BODY_FONT}, {VOICE_TONE}, {BRAND_NAME_SLUG}. The user must see only the finished prompt — no {...} markers left over.
Then present to the user, framed exactly like this:
Here's a prompt I've put together for your Meta Post template. Copy and paste it into Claude Design:
- Open https://claude.ai/design in your browser
- Create a new project (name it whatever you like — e.g. "Acme Meta Post Template")
- Paste the prompt below into Claude Design's chat:
"I need a configurable Instagram + Facebook single-image feed post template for {BRAND_NAME}. 4:5 portrait, 1080×1350, ONE slide. Build it as a React + Babel app rendered from a single index.html (use Babel standalone via CDN — no build step). The agent will modify copy programmatically per post, so structure matters more than specific sample copy. The brand uses {PRIMARY_HEX} as primary, {SECONDARY_HEX} as secondary, {ACCENT_HEX} as accent. Headings are set in {HEADING_FONT}, body in {BODY_FONT}. The aesthetic is {VOICE_TONE}. Include the {BRAND_NAME} brand logo subtly in a corner of the slide and the handle as a small footer (e.g. @{BRAND_NAME_SLUG}). This same image will be published to both Instagram feed and Facebook feed — design accordingly (no platform-specific UI chrome). Three direction styles (A / B / C) sharing the same EDITMODE keys but different visual treatments — the agent picks a direction per post via the `direction` key: - Direction A — Hero Visual (image-led): full-bleed background photo with a theme overlay, oversized emotional headline as the centerpiece, eyebrow chip above, minimal body underneath, accent-color CTA pill at bottom. Best for brand campaigns, product launches, scroll-stopping visuals. - Direction B — Quote Card (editorial): large quote in serif/italic as the centerpiece, attribution line beneath, optional small headshot circle. Cream / light or branded background. Best for testimonials, founder voice, press quotes. - Direction C — Listicle Teaser (educational): eyebrow + headline at top, then 3–5 numbered or bulleted points down the slide (b1–b5), CTA at bottom. Best for value-led / educational posts (\"5 ways to...\", \"3 reasons why...\"). CRITICAL CONTRACT — wrap all editable copy in a single JSON object inside index.html, marked with /*EDITMODE-BEGIN*/ and /*EDITMODE-END*/ comment markers, using these EXACT key names:window.__TWEAK_DEFAULTS = /EDITMODE-BEGIN/{ "direction": "A", "accent": "purple", "_template_theme": "dark", "handle": "@{BRAND_NAME_SLUG}", "hashtag": "#YourTag", "eyebrow": "sample eyebrow", "headline": "sample IG/FB headline that stops the scroll", "body": "sample supporting body line, ~1 sentence max for IG/FB.", "quote": "sample quote that fits a quote-card treatment.", "attribution": "Jane Doe, Customer", "bg_image": "", "bg_image_position": "center", "bg_image_fit": "cover", "headshot_image": "", "b1": "First bullet point", "b2": "Second bullet point", "b3": "Third bullet point", "b4": "Fourth bullet point (optional)", "b5": "Fifth bullet point (optional)", "cta_text": "Tap the link in bio", "cta_button": "Learn more" }/EDITMODE-END/;
Use these EXACT key names — the agent substitutes copy by parsing the JSON between the markers and writing back. The `direction` key (A / B / C) selects which visual treatment renders. Unused keys for a given direction (e.g. quote/attribution when direction='A', or b1–b5 when direction='B') should be ignored gracefully, not shown. The agent may also pass fewer than 5 bullets to direction C — design the layout so it still looks balanced with only 3 bullets. THEME INDICATOR — set `_template_theme` to `\"dark\"` or `\"light\"` to match direction A (which receives the full-bleed photo). The template uses it to pick the native overlay color (dark = rgba(0,0,0,0.4), light = rgba(255,255,255,0.5)). fb.ai surfaces it as `manifest.theme` but composites nothing — the template owns the overlay. EXPORT MARKERS — fb.ai's renderer screenshots the active direction by `data-export-id`, capturing it at full 1080×1350. Each direction artboard's outer `<div>` (the full-size 1:1 slide) carries a **static string literal** `data-export-id`, e.g. `data-export-id=\"A · Hero Visual\"`, `data-export-id=\"B · Quote Card\"`, `data-export-id=\"C · Listicle\"` (name them to match the direction label). Never a JSX expression or a quoted template literal (`data-export-id=\"${dir}\"`) — the uploader scans statically. All three artboards stay in the DOM; gate the inactive directions with `visibility:hidden`, `opacity:0`, or an off-screen offset — **not** `display:none`, which yields a blank capture if that direction is the one exported. IMAGE SLOTS — direction A takes a full-bleed user photo via `bg_image` / `bg_image_position` / `bg_image_fit`; direction B takes an optional circular `headshot_image`; direction C is design-only. Render them NATIVELY per the native-image contract above: full-bleed `<img>` at zIndex:0 (objectFit/objectPosition from the companion keys), theme overlay div at zIndex:1, content wrapper at position:relative; zIndex:2; when `bg_image` is empty show the gradient/placeholder. Direction B must look balanced when `headshot_image` is empty. The brand logo is bundled at `assets/logo.png` and referenced directly — never in EDITMODE. No base64 data URIs, no bundled photo/demo assets, no unused fonts. SAFE ZONES — Instagram doesn't crop 4:5 feed images, but the profile-grid recrop (~34px side trim) can clip edge content. Keep critical text and logo at least ~80px from the left and right edges. Facebook's feed renders 4:5 fully but the bottom ~60px can collide with reaction/comment chrome on some clients — keep critical content out of the bottom 60px. Use the brand colors and fonts above. The template's sample copy will be replaced at runtime — don't worry about it being final."Iterate with Claude inside Claude Design until all three directions (A/B/C) look on-brand. Don't worry about the sample copy; it gets replaced per post at runtime. Once the layout looks good, let me know and we'll move to Step B (export).
Wait for the user to confirm they're happy with the template before exporting. Then run the shared install flow above (Step B export → Step C upload to the fb.ai dashboard → Step D verify via fivebucks_list_templates). The meta-post manifest should expose direction, _template_theme, the copy keys (eyebrow/headline/body/quote/attribution/b1–b5/cta_text/cta_button), and the bg_image trio (+ optional headshot_image). If the user skips this template, IG/FB single-image posts fall back to the Gemini + Pillow image-path.
Do not proceed to Step 5 until 4b, 4c, and 4d have been addressed — installed or explicitly skipped. Either outcome is acceptable; Step 9c will record the actual state in CLAUDE.md.
Step 4d — Media Library (OPTIONAL — recommended)
Upload brand photos and visuals to the fb.ai media library so content-generator and creative-designer can pull on-brand images into posts at runtime.
Ask the user:
Want to upload brand photos to the fb.ai media library? This lets skills pick on-brand visuals automatically when generating posts. You can skip and add media later.
If the user says skip, acknowledge and move on to Step 5. If yes, walk them through it:
Go to https://www.fivebucks.ai/dashboard/api-keys and sign in — generate your API Key and save it somewhere safe. (Skip if you have already generated your API Key.)
Create one folder per template type you have set up (Step 4c), named exactly as follows —
content-generatoruses these names to automatically match photos to the right template at runtime:
- LinkedIn Post — photos for LinkedIn single-image posts
- Meta Story — photos for Instagram / Facebook Stories
- Meta Carousel — photos for Instagram / Facebook Carousels
- Meta Post — photos for Instagram / Facebook single-image posts
Only create folders for template types you actually have on fb.ai. Then click Upload Media inside each folder to add your photos.
Confirm once the user is done, then move on to Step 5.
Step 5 — Research & Context Generation
This is the big one. Step 5 generates 8 separate context files in brands/{brand}/ — product.md, competitors.md, funnel.md, plus the v2.4.0 additions: sales.md, customer-success.md, finance.md, investors.md, operations.md. Each file feeds different downstream skills. The first 4 are extracted from your website automatically by /link-skills:research-strategy; the new v2.4.0 files are operational config you provide directly (only you know your ICP filters, payment terms, investor list).
Expect 30–60 minutes total depending on how many optional sub-steps apply. Each sub-step is independently skippable — for example, investors.md only matters if you've raised outside funding. Smart defaults are offered everywhere; you can usually just confirm and move on.
Now that brand.md and audience.md exist, run /link-skills:research-strategy to fill in the remaining context files. The research skill will:
- Read brand.md + audience.md for context
- Research competitors automatically (no need to ask the user)
- Analyze market positioning, strengths, weaknesses
- Keyword research is skipped at this stage — DataforSEO keys are not entered until Step 7. After completing Step 7, the user can re-run
/link-skills:research-strategyto add keyword data to the strategy outputs.
Read the research output from outputs/{brand}/strategy/ and use it to generate:
brands/{brand}/product.md
# {Brand Name} — Product
## Overview
{1-2 sentence summary}
## Features
{extracted from website in Step 4}
## Pricing
{extracted from pricing page or "Ask user"}
## Differentiators
{what makes this product unique — informed by competitive research}
brands/{brand}/competitors.md
Generated from research-strategy output. No need to ask the user — competitors are discovered automatically. The monitor_urls, track_pages, and exec_team fields below are added per-competitor to support the competitor-monitor skill (weekly diff of pricing pages, blog index, exec hires, etc.):
# {Brand Name} — Competitive Landscape
## {Competitor 1}
- **URL:** {url}
- **Positioning:** {how they position themselves}
- **Strengths:** {what they do well}
- **Weaknesses:** {where they fall short}
- **Counter-messaging:** {how to position against them}
- **monitor_urls:**
- homepage: {https://competitor.com/}
- pricing: {https://competitor.com/pricing}
- blog: {https://competitor.com/blog}
- careers: {https://competitor.com/careers}
- changelog: {https://competitor.com/changelog or release-notes URL — omit if absent}
- **track_pages:**
- pricing: price changes, new tier introductions, removed tiers
- homepage: hero copy / positioning shifts, new product mentions
- blog: new launches, thought-leadership themes
- careers: new exec hires (esp. C-level / VP roles), team-size growth signals
- changelog: new feature releases, deprecations
- **exec_team:**
- {Name, Title} — {LinkedIn URL}
- {Name, Title} — {LinkedIn URL}
## {Competitor 2}
...
Populate monitor_urls automatically from research-strategy output where possible (homepage and pricing are usually known; blog / careers / changelog can be inferred from the homepage navbar). Populate exec_team from the competitor's About / Team / Leadership page if research-strategy surfaced it; otherwise leave with a TBD — fill manually from LinkedIn placeholder. The competitor-monitor skill reads these fields directly — no user prompt at this stage.
brands/{brand}/funnel.md
Step A — Ask the user about their funnel:
What does your conversion funnel look like? What happens after someone clicks your ad? (e.g. "website visit → WhatsApp chat → close deal" or "visit → trial signup → paid conversion")
Step B — Discover actual GA4 events:
Windsor.ai is set up in Step 7, so at this point it is likely not connected yet. Ask the user for their GA4 key event / custom event names directly. If they don't know, leave event names as TBD — they will be discovered and filled in during Step 8 (Connection Validation) once Windsor.ai is connected.
If Windsor.ai happens to already be connected (e.g. returning user), use Windsor MCP tool get_fields with source: "googleanalytics4" to discover actual events now. Look for event names that match the user's described funnel actions (e.g. click_whatsapp, click_email, schedule_call, signup_form_submit, trial_activated, purchase).
Show the user the relevant events found and confirm the mapping:
I found these key events / custom events that match your funnel: [list]. Can you confirm which event maps to each step?
Step C — Generate funnel.md with the confirmed mapping:
# {Brand Name} — Conversion Funnel
## Funnel Type
{e.g. "Lead gen", "SaaS trial", "E-commerce", "Service inquiry"}
## Stages (Google Ads)
| Stage | GA4 Event / Source | Benchmark | Status Threshold |
|---|---|---|---|
| Impressions | Google Ads | — | — |
| Clicks | Google Ads | CTR >{X}% | 🟢 >{X}% / 🟡 {Y}-{X}% / 🔴 <{Y}% |
| GA4 Sessions | GA4: session_source_medium = google / cpc | {X}-{Y}% of clicks | 🟢/🟡/🔴 |
| {Conversion Action 1} | GA4: {event_name} | {X}% of sessions | 🟢/🟡/🔴 |
| {Conversion Action 2} | GA4: {event_name} | {X}% of {previous stage} | 🟢/🟡/🔴 |
{...add as many stages as the client's funnel has}
## Stages (Meta Ads)
| Stage | GA4 Event / Source | Benchmark | Status Threshold |
|---|---|---|---|
| Impressions | Meta Ads | — | — |
| Clicks | Meta Ads | CTR >{X}% | 🟢/🟡/🔴 |
| LP Views | Meta Ads | >{X}% of clicks | 🟢/🟡/🔴 |
| GA4 Sessions | GA4: session_source_medium = meta / paid_social | {X}-{Y}% of clicks | 🟢/🟡/🔴 |
| {Conversion Action 1} | GA4: {event_name} | {X}% of sessions | 🟢/🟡/🔴 |
| {Conversion Action 2} | GA4: {event_name} | {X}% of {previous stage} | 🟢/🟡/🔴 |
{...same conversion actions, different source filter}
## Cost Benchmarks
| Metric | Target ({currency from brand.md Locale}) |
|---|---|
| Cost per {Conversion Action 1} | {X} |
| Cost per {Conversion Action 2} | {X} |
## Notes
- Benchmarks are initial estimates. Update after 2-4 weeks of data.
- If GA4 events are TBD, the digital-marketing-analyst will skip funnel stages without mapped events.
Each client's funnel is unique — the digital-marketing-analyst reads this file and builds the email funnel table dynamically from whatever stages are defined here.
Do not proceed until product.md, competitors.md, and funnel.md are generated and the user has confirmed the funnel mapping.
Step 5g — Sales context
The first new v2.4.0 file.
sales.mdpowers your sales engine —apollo-lead-prospectorreads ICP filters from it to query Apollo,outreach-sequenceruses the sender persona and sequence templates to drive cold email loops,proposal-generatoruses the proposal defaults when packaging deals, and (v2.16.0)gig-prospectorreads the Inbound Job Filters (Step H) to discover freelance jobs the brand can bid on. Without this file, those skills can't run on this brand at all.Expect ~7 questions over 5–10 minutes. Step B (ICP filters) repeats once per persona, so it scales with how many personas you defined in Step 4. Most prompts have smart defaults — Apollo people-search filters can be derived from
audience.md, daily quota defaults to 20/persona, sequence shape defaults to 5-touch over 14 days unless you specify otherwise.
Generate brands/{brand}/sales.md. Unlike product.md / competitors.md (which are extracted from website research), this file is operational config — prompt the user for the values directly. The file is read by apollo-lead-prospector, outreach-sequencer, and proposal-generator.
Read existing context first — do not ask the user what's already in another brand file. Before any prompt below, read these files and pre-populate draft answers. Then show the user the drafted block for each step and ask "Confirm or edit?" — never ask a question from scratch when the source file already has the answer:
| File | Field to read | Step it pre-fills |
|---|---|---|
brand.md |
Founder name, title, voice samples, signature block (if present) | Step A (Sender Persona — Name / Title / Signature draft), Step E (Sequence Templates — tone matching) |
audience.md |
Personas + their job titles, industries, company sizes, geos, pain points | Step B (ICP Filters — pre-drafted per persona), Step E (Sequence Templates — pain-point hooks per persona) |
competitors.md |
Competitor URLs | Step C (Disqualification — derive competitor_domains blocklist from URLs; auto-set "skip if competitor employee" toggle to yes) |
product.md |
Pricing section tier names | Step G (Proposal Defaults — Default tier per persona) |
product.md |
Overview / Features / Differentiators (services + tools the brand sells) | Step H (Inbound Job Filters — drafts the Search Keywords from what the brand actually delivers; never hardcoded) |
brand.md |
Locale / geography | Step H (Inbound Job Filters — proposes the default Markets to monitor) |
If a source file is empty or absent, fall back to asking the user from scratch for that field only — don't block the whole sub-step.
Step A — Sender Persona:
Who is the human sender that outbound emails should appear to come from? I need:
- Name
- Title
- Email signature block (paste the exact block you'd want appended to outbound emails)
- Photo URL (LinkedIn or website headshot — used in proposals and on the email signature)
- Booking URL — your public Calendly scheduling link (e.g.
https://calendly.com/you/intro-call). Used as the standing CTA link in outreach emails when the per-prospect single-use Calendly tool isn't available, incustomer-onboarderwelcome emails when the kickoff event-type lookup fails, and to resolve the{link}placeholder in Reply Routing below.
Step B — ICP Filters (per persona):
Re-read brands/{brand}/audience.md to enumerate the personas defined in Step 4. For each persona, ask the user to confirm Apollo people-search filters that will return matching leads:
- Titles — list of job titles to match (e.g.
["VP of Marketing", "Head of Growth", "Marketing Director"]) - Industries — list of industries (e.g.
["SaaS", "B2B Software"]) - Company size band — employee-count range (e.g.
11-50,51-200,201-1000) - Geography — countries / regions (e.g.
["United States", "Singapore"]) - Tech stack (optional) — required tech the company must use (e.g.
["HubSpot", "Salesforce"])
Step C — Disqualification Rules:
What leads should we automatically skip? Common entries:
- Blocklist domains (existing customers, partners, your own domain)
- "Skip if already a customer" (toggle yes/no)
- "Skip if competitor employee" (toggle yes/no — uses
competitors.mdURLs to derive domains)- Any other exclusion rules
Step D — Daily Quota:
How many new prospects should be added per day per persona? (Default: 20 per persona unless you state otherwise.)
Step E — Sequence Templates:
For each persona, capture the touch shape. If the user doesn't specify, infer a default from the persona's pain points in audience.md (e.g. for a senior buyer: "5-touch over 14 days: Day 0 email / Day 3 email / Day 7 email / Day 10 LinkedIn DM / Day 14 final email").
Step F — Reply Routing:
Default rules — confirm with the user, but only ask if they want to deviate:
- "interested" → book a meeting (Calendly link)
- "not now" → 90-day nurture loop
- "wrong person" → ask for referral, then add original contact to disqualified list
- "unsubscribe" → permanent skip, never contact again
Step G — Proposal Defaults:
When
proposal-generatorwrites a proposal, what defaults should it use?
- Payment terms (e.g.
Net 30,50% upfront / 50% on delivery)- Validity period (default: 30 days)
- Default tier per persona (read tier names from
brands/{brand}/product.mdPricing section)- Upsell rules (e.g. "always offer annual billing with 15% discount", "include onboarding add-on for Enterprise")
Step H — Inbound Job Filters (gig-prospector):
gig-prospector(v2.16.0) is the inbound counterpart toapollo-lead-prospector: instead of sourcing people to email, it scans freelance marketplaces every day for open jobs the brand can bid on, scores them for fit, and drops matches into a Notion opportunities DB. This block configures where it looks and what it looks for. Skip it if the brand does not pursue marketplace/freelance work —gig-prospectorsimply won't run for this brand.
Capture four things. Pre-draft from context first (per the table above) — only ask the user to confirm or edit:
- Markets — which markets to monitor. Pre-propose from
brand.mdLocale/geography; offer the full set:Singapore,Indonesia,Malaysia,Thailand,Australia,Global-Remote. Multi-select.Which markets should I watch for freelance jobs? (Singapore / Indonesia / Malaysia / Thailand / Australia / Global-Remote — pick any.)
- Platforms + account status — which marketplaces to search, and whether the brand has an account on each (account status decides whether
gig-prospectorcan read login-walled listings vs public pages only). Propose the platforms that serve the chosen markets:
Record each asFor each platform you want me to scan, tell me if you already have an account there (so I can reach listings behind a login). Recommended set by market: Upwork (global), Freelancer.com (global + SG/MY/ID/AU), Projects.co.id / Sribu (Indonesia), Fastwork (Thailand/Indonesia), Freelancing.my (Malaysia), plus PeoplePerHour (global, secondary) and Jobbers.io (SG/MY, lower-confidence). Airtasker (Australia) only if you do task/services work — it's low-yield for tech/automation. Singapore has no dedicated project marketplace (use the global platforms). Skip gig-catalog sites (Fiverr) and employment boards (JobStreet/Indeed) — they aren't project-bidding marketplaces. Add any others you use.
platform — have account/platform — no account. If they have a Freelancer.com account and want the deterministic API source, note that an optionalFREELANCER_OAUTH_TOKENcan be added later (Step 7b vault) to scan it without browser scraping. - Search Keywords — derive these from
product.md, do not hardcode. Read the Overview / Features / Differentiators and draft the concrete services + tools the brand sells (e.g. the platforms it builds on, "integration", "workflow", the apps it connects), then show the drafted list and ask the user to confirm/trim/add. These become the per-platform search terms. - Budget Floor, Exclusions, Daily Cap — minimum acceptable budget/rate (optional), keywords/client-types/geos to skip, and max opportunities to add per day (default 30).
Account & login note:
gig-prospectorreads listings through Claude in Chrome (the user's own authenticated browser — it beats Cloudflare and reads behind logins) and pauses for the user to sign in if a login wall appears. So "have account" platforms get full coverage; "no account" platforms get public-page coverage only. This is configuration, not a connection step — no OAuth here.
Save as brands/{brand}/sales.md with the following sections:
# {Brand Name} — Sales Operations
## Sender Persona
- Name: {...}
- Title: {...}
- Signature:
{paste signature block}
- Photo URL: {...}
- Booking URL: {https://calendly.com/...}
## ICP Filters
### {persona-slug-1}
- Titles: [...]
- Industries: [...]
- Company size: {band}
- Geography: [...]
- Tech stack: [...] # optional
### {persona-slug-2}
...
## Disqualification Rules
- Blocklist domains: [...]
- Skip if already customer: yes/no
- Skip if competitor employee: yes/no
- Other: ...
## Daily Quota
- {persona-slug-1}: {N} prospects/day
- {persona-slug-2}: {N} prospects/day
## Sequence Templates
### {persona-slug-1}
{e.g. 5-touch over 14 days: Day 0 email / Day 3 email / Day 7 email / Day 10 LinkedIn / Day 14 final email}
### {persona-slug-2}
...
## Reply Routing
- interested → book meeting (Calendly: {Booking URL from Sender Persona above})
- not now → 90-day nurture
- wrong person → ask for referral, disqualify original
- unsubscribe → permanent skip
## Proposal Defaults
- Payment terms: {Net 30 / etc.}
- Validity period: {30 days}
- Default tier per persona:
- {persona-slug-1}: {tier name from product.md}
- {persona-slug-2}: {tier name}
- Upsell rules: ...
## Inbound Job Filters
# Read by gig-prospector (v2.16.0). Omit this whole section if the brand does not pursue freelance/marketplace work.
- Markets: [Singapore | Indonesia | Malaysia | Thailand | Australia | Global-Remote] # any subset
- Platforms: # each tagged "have account" / "no account"
- {platform-1}: have account | no account
- {platform-2}: have account | no account
- Search Keywords: [...] # DERIVED from product.md (services + tools the brand sells); never hardcoded
- Budget Floor: {e.g. $500 fixed / $25/hr — or "none"}
- Exclusions: [...] # keywords / client types / geographies to skip
- Daily Cap: {N} # max opportunities added per day (default 30)
Step 5h — Customer Success context
customer-success.mdpowers two retention skills —customer-onboarderwalks new customers through your milestone playbook (welcome email, kickoff, milestone tracking), andchurn-predictorscores active customers daily against the health-score formula you define here, alerting on at-risk transitions. If you skip this file, those 2 skills can't run.Expect ~6 questions over 8–12 minutes. The biggest input is Step B (Onboarding Milestones) — you'll define 5–7 milestones per plan tier, so 3 tiers means 15–21 milestones total. The health-score weights (Step D) and intervention playbook (Step E) come with strong defaults; many founders just confirm those.
Generate brands/{brand}/customer-success.md. Read by customer-onboarder and churn-predictor. Operational config — prompt the user.
Read existing context first — do not ask the user what's already in another brand file. Pre-populate draft answers from these sources, then ask "Confirm or edit?" per block:
| File | Field to read | Step it pre-fills |
|---|---|---|
product.md |
Pricing section tier names + descriptions | Step A (Plan Tiers — pre-fill tier list and short descriptions) |
audience.md |
Personas per tier (if mapped) + their key pain points | Step B (Onboarding Milestones — tailor milestone wording per persona's pain), Step E (Intervention Playbook — match outreach style to persona) |
funnel.md |
GA4 conversion events / activation events | Step B (Onboarding Milestones — propose trigger event names from real funnel events instead of asking the user to invent them) |
brand.md |
Voice samples, tone | Step E (Intervention Playbook — re-engagement email wording sounds on-brand), Step F (NPS check-in copy) |
sales.md (if Step 5g already ran) |
Sender Persona signature | Step E (CSM personal outreach — use same signature unless user overrides) |
If a source is missing, fall back to asking the user from scratch for just that field.
Step A — Plan Tiers:
Re-read brands/{brand}/product.md Pricing section to enumerate tiers (e.g. Free / Pro / Enterprise). Confirm the tier list with the user.
Step B — Onboarding Milestones per tier:
For each tier, prompt the user for 5-7 milestones. Each milestone has: name, trigger event (the system event that fires it), expected days from signup.
For your {tier} plan, what are the onboarding milestones a healthy customer hits? Example for a Pro tier: signup (Day 0) → setup wizard complete (Day 0-1) → first asset created (Day 1-3) → first-week active (Day 7) → 30-day retention check (Day 30).
Step C — Kickoff Agenda (Enterprise / high-touch tiers only):
For tiers that include a kickoff call, prompt for agenda. Default: intro / use-case discovery / product walkthrough / next steps. Skip for tiers without a kickoff (e.g. Free, Pro self-serve).
Step D — Health Score Weights:
Default formula (confirm with user, override only if they want to deviate):
- 40% feature adoption (count of key features used in last 30 days vs total)
- 30% login frequency (sessions per week)
- 20% support volume (inverse — fewer tickets = healthier)
- 10% subscription state (active / past-due / cancelled)
Capture thresholds for each component (e.g. "feature adoption ≥ 60% = healthy, 30-60% = watch, <30% = at-risk").
Step E — At-Risk Intervention Playbook:
Define what triggers each band and what to do at each:
- Healthy (score ≥80): no action; quarterly check-in
- Watch (60-79): automated re-engagement email with usage tip
- At-Risk (40-59): CSM personal outreach within 48h
- Critical (<40): founder/CSM escalation, win-back call within 24h
Step F — Day-X Check-in Cadence:
When to send NPS / satisfaction surveys. Default: Day 7, Day 30, Day 90, then quarterly. Confirm with user.
Save as brands/{brand}/customer-success.md with sections:
# {Brand Name} — Customer Success Operations
## Plan Tiers
- {Tier 1}: {short description}
- {Tier 2}: ...
## Onboarding Milestones
### {Tier 1}
1. {Milestone name} — trigger: {event}, expected: Day {N}
2. ...
### {Tier 2}
...
## Kickoff Agenda
({tiers that include kickoff})
1. {agenda item}
2. ...
## Health Score Weights
- Feature adoption: 40% (threshold: ≥60% healthy)
- Login frequency: 30% (threshold: ≥3 sessions/week healthy)
- Support volume: 20% (threshold: ≤2 tickets/month healthy)
- Subscription state: 10%
## At-Risk Intervention Playbook
- Healthy (≥80): quarterly check-in
- Watch (60-79): automated re-engagement email
- At-Risk (40-59): CSM personal outreach within 48h
- Critical (<40): founder escalation, win-back call within 24h
## Day-X Check-in Cadence
- Day 7: NPS pulse
- Day 30: NPS + product-fit survey
- Day 90: NPS + renewal-intent
- Quarterly thereafter
Step 5i — Finance context
finance.mdis the brain behind your back-office automation.invoice-collectorreads the escalation tone ladder to chase overdue invoices day by day;financial-reporterreads the KPIs and alert thresholds to publish your monthly P&L summary and Slack-alert if runway dips below your floor. Without this file, those 2 skills can't run.Expect ~8 questions over 10–15 minutes. This sub-step has the most jargon — terms like "revenue recognition method", "runway calc window", "alert thresholds". Don't sweat it; defaults are SaaS-standard (Net 30 / 6-month runway floor / 3-month avg burn / cash-basis reporting). The most important input is Step B (Escalation Tone Ladder) — the agent drafts your D+1 / D+7 / D+14 / D+30 reminder copy in your brand voice and asks you to review each. Budget extra time there.
Generate brands/{brand}/finance.md. Read by invoice-collector and financial-reporter. Operational config — prompt the user; pull voice from brand.md for the escalation ladder so reminders sound on-brand.
Read existing context first — do not ask the user what's already in another brand file. Pre-populate draft answers, then ask "Confirm or edit?" per block:
| File | Field to read | Step it pre-fills |
|---|---|---|
brand.md |
Voice samples, tone | Step B (Escalation Tone Ladder — draft D+1 / D+7 / D+14 / D+30 wording in brand voice; user reviews not invents) |
brand.md |
## Locale section (Currency / Timezone) |
Step F (Alert Thresholds — use the right currency symbol; show AR threshold in local currency) |
product.md |
Pricing section (subscription tiers vs one-off / project-based) | Step E (KPIs — auto-include MRR/ARR for subscription pricing; swap for project-margin / cash-collected for services), Step H (Revenue Recognition — default cash for one-off / SMB, default accrual if Pricing implies recurring revenue) |
sales.md (if Step 5g already ran) |
Sender Persona signature | Step B (Escalation Ladder — append signature to D+30 final notice for human-feel) |
If a source is missing, fall back to asking the user from scratch for just that field.
Step A — Payment Terms:
What's your default invoice payment term? (Default: Net 30. Common alternatives: Net 7, Net 14, Net 60, "due on receipt", "50% upfront / 50% on completion".)
Step B — Escalation Tone Ladder:
Compose exact wording for D+1, D+7, D+14, D+30 reminders using the voice in brand.md. Sample structure:
- D+1 — friendly nudge: "Hi {first_name}, hope this finds you well. Just a quick reminder that invoice {#} for {amount} is due. Here's the payment link: {link}. Let me know if you have any questions!"
- D+7 — direct follow-up: "Following up on invoice {#} — it was due {N} days ago. Could you let me know when we can expect payment, or flag any issues?"
- D+14 — firm with payment link: "Invoice {#} is now 14 days overdue. To avoid service disruption, please pay at {link} or reach out to discuss."
- D+30 — final notice: "Final notice on invoice {#}. If unpaid by {date}, we'll need to {pause service / refer to collections / escalate to founder}. Please reply today to resolve."
Adjust each tier to brand voice (e.g. a casual SMB brand softens D+30; a B2B enterprise brand stays formal throughout).
Step C — Retry Intervals:
How often should the agent retry between rungs? (Default: every 7 days — D+1 → D+7 → D+14 → D+30.)
Step D — Dispute Handling:
When a customer flags a dispute, the agent must pause chasing. Default: "If the customer replies with words like 'dispute', 'incorrect', 'wrong amount', or 'cancel' → mark dispute=true in Xero metadata (or Notion if Xero unavailable) and skip that invoice from all future sweeps. Notify the founder via Slack."
Step E — KPIs to Highlight:
Default list (confirm with user; SaaS brands keep MRR/ARR, services brands swap for project margin):
- MRR (monthly recurring revenue)
- ARR (annual recurring revenue)
- Gross margin
- Runway (cash / monthly burn)
- Top revenue movers (week-over-week)
- Top expense movers (week-over-week)
Step F — Alert Thresholds:
When financial-reporter should wake the founder up (Slack DM):
- Runway < N months (default: 6)
- MRR drop > X% MoM (default: 5%)
- Single expense category up > Y% MoM (default: 25%)
- AR > Z days outstanding total (default: $50k or 30 days, whichever first)
Step G — Runway Calc Method:
How should the agent compute runway? (Default: cash on hand / monthly burn averaged over the last 3 months. Alternatives: trailing 6 months, single most-recent month.)
Step H — Revenue Recognition Rules:
Cash basis or accrual? (Default: cash for SMB / bootstrapped brands; accrual for VC-backed brands required to report investor-grade financials.)
Save as brands/{brand}/finance.md with sections matching each block:
# {Brand Name} — Finance Operations
## Payment Terms
{Net 30 / etc.}
## Escalation Tone Ladder
### D+1 (friendly nudge)
{wording}
### D+7 (direct follow-up)
{wording}
### D+14 (firm with payment link)
{wording}
### D+30 (final notice)
{wording}
## Retry Intervals
Every {N} days between rungs.
## Dispute Handling
{rules — keyword triggers, pause logic, who to notify}
## KPIs to Highlight
- MRR / ARR / gross margin / runway / top movers / ...
## Alert Thresholds
- Runway < {N} months → Slack alert
- MRR drop > {X}% MoM → Slack alert
- Expense category up > {Y}% MoM → Slack alert
- AR > {Z} days outstanding → Slack alert
## Runway Calc Method
{cash / N-month-avg burn}
## Revenue Recognition Rules
{cash / accrual}
Step 5j — Investors context (fundraised brands only)
investors.mdpowersinvestor-update-writer— a monthly cron skill that combines your Xero financials, PostHog product KPIs, and CRM wins into a draft email that lands in your Gmail Drafts folder for review and send. It mirrors your past update style by reading prior monthly updates you paste in.Skip this entirely if you haven't raised outside funding — the consent gate below makes that explicit. If you have raised, expect ~5 questions over 15–20 minutes; the heaviest input is Step E (Prior Updates Archive), where you'll paste your last 3-6 monthly updates. That's a one-time copy-paste from your email or wherever you store them. Without those samples, the drafted updates will sound generic.
Step 0 — consent gate
Ask the user:
Have you raised outside funding (angels, VCs, seed/Series A+) and want Claude to draft your monthly investor updates?
If no, skip this sub-step entirely — investor-update-writer won't be used. If yes, generate brands/{brand}/investors.md:
Read existing context first — do not ask the user what's already in another brand file. Pre-populate draft answers, then ask "Confirm or edit?" per block:
| File | Field to read | Step it pre-fills |
|---|---|---|
brand.md |
Voice samples, founder tone | Step B (Founder Voice — fall back to brand voice if user pastes <2 prior updates), Step C (Sections to Include — drafting tone) |
finance.md (if Step 5i already ran) |
KPIs to Highlight + Alert Thresholds | Step C (Sections to Include — auto-add the same KPIs to the monthly update, so the dashboard alert thresholds and investor narrative use one source of truth) |
sales.md (if Step 5g already ran) |
Sender Persona name + signature | Step E (drafting attribution — updates send under the same sender as outbound) |
customer-success.md (if Step 5h already ran) |
Health-score bands | Step C (Wins / Lowlights drafting — auto-call out customers transitioning across health bands) |
If a source is missing, fall back to asking the user from scratch for just that field.
Step A — Investor List:
Accept CSV paste. Each row has: name, firm, email, role (lead / follow / angel / advisor), preferred update frequency (monthly / quarterly).
Paste your investor list as CSV (one row per investor). Columns: name, firm, email, role, frequency. Example:
Jane Smith,Acme Ventures,jane@acme.vc,lead,monthly John Doe,Angel,john@example.com,angel,quarterly
Step B — Founder Voice:
Paste 2 prior monthly investor updates verbatim (or 1 if only one exists). The skill uses these as voice samples — your asks, your wins framing, your phrasing patterns. Without this, drafts will sound generic.
Step C — Sections to Include:
Default: TL;DR / KPIs / Wins / Lowlights / Asks / Hires. Confirm with user; allow override (e.g. some founders skip Lowlights for early-stage updates).
Step D — Sections to OMIT:
Default: never share specific customer names without consent, internal team conflicts, unconfirmed M&A discussions, churned-customer details. Capture any brand-specific additions.
Step E — Prior Updates Archive:
Paste your last 3-6 monthly updates (most recent first). The skill reads these to avoid repeating asks/wins and to maintain narrative continuity.
Save as brands/{brand}/investors.md with sections:
# {Brand Name} — Investor Communications
## Investor List
| Name | Firm | Email | Role | Frequency |
|---|---|---|---|---|
| {...} | {...} | {...} | {lead/follow/angel/advisor} | {monthly/quarterly} |
## Founder Voice (samples)
### Sample 1 — {month} {year}
{paste verbatim}
### Sample 2 — {month} {year}
{paste verbatim}
## Sections to Include
- TL;DR
- KPIs (MRR, ARR, runway, key funnel metrics from finance.md)
- Wins
- Lowlights
- Asks (intros, hires, advice)
- Hires (new joiners)
## Sections to OMIT
- Specific customer names without consent
- Internal team conflicts
- Unconfirmed M&A discussions
- Churned customer details
## Prior Updates Archive
### {Month Year}
{paste full update}
### {Month Year}
{paste full update}
...
Step 5k — Operations context (optional)
operations.mdpowersmeeting-analyzer— a skill that processes meeting transcripts (1:1s, standups, client calls, board meetings) into structured action items + decisions, routes each owner to the right Slack handle, and drafts follow-up emails for client meetings. The routing rules in this file are how the skill knows that "engineering tasks → Tech Lead", "marketing tasks → CMO", and so on.This file is fully optional.
meeting-analyzerworks without it — action items get marked "Unassigned" and you reassign manually in Notion. Skip if you're a solo founder or don't run formal meetings with multiple owners. If you do opt in, expect ~3 questions over 5 minutes.
Ask the user:
Do you run regular meetings (1:1s, standups, client calls, board meetings) and want Claude to process transcripts — extract action items, route them to owners, and archive notes?
If no, skip this sub-step entirely. meeting-analyzer works without it (degrades to "unassigned" owners on action items). If yes, generate brands/{brand}/operations.md:
Read existing context first — do not ask the user what's already in another brand file. Pre-populate draft answers, then ask "Confirm or edit?" per block:
| File | Field to read | Step it pre-fills |
|---|---|---|
sales.md (if Step 5g already ran) |
Sender Persona name | Step A (Action Item Routing — pre-fill Sales tasks owner with the sender persona) |
customer-success.md (if Step 5h already ran) |
CSM mentions in Step E intervention playbook | Step A (Action Item Routing — pre-fill Customer issues owner with the named CSM if present) |
finance.md (if Step 5i already ran) |
Founder/CFO mentions in Step F alert thresholds | Step A (Action Item Routing — pre-fill Finance tasks owner with the alert recipient) |
brand.md |
Founder name | Step C (Default Owners — fallback owner is the founder if no other owners are explicitly defined for a category) |
If a source is missing, fall back to asking the user from scratch for that owner only.
Step A — Action Item Routing:
For each meeting type, capture who owns each action category. Example mapping:
- Engineering tasks → Tech Lead (Slack handle:
@alice) - Marketing tasks → CMO (Slack handle:
@bob) - Sales tasks → Sales Lead (Slack handle:
@carol) - Finance tasks → Founder/CFO (Slack handle:
@dave) - Customer issues → CS Lead (Slack handle:
@eve)
Step B — Meeting Types:
Each meeting type can have a different output template. Default types:
- 1:1 — agenda items, blockers, career-development notes
- standup — yesterday/today/blockers per attendee
- client — agenda, decisions, action items, customer sentiment
- sales — qualification answers, objections, next-step commitments
- board — KPI review, strategic decisions, asks
Step C — Default Owners:
Fallback when an action item has no clear owner (e.g. transcript is ambiguous). Default: route to the meeting organizer's Slack handle, with a 24h "claim or reassign" window.
Save as brands/{brand}/operations.md:
# {Brand Name} — Operations
## Action Item Routing
| Category | Owner | Slack Handle |
|---|---|---|
| Engineering | {Tech Lead name} | @{handle} |
| Marketing | {CMO name} | @{handle} |
| Sales | {Sales Lead name} | @{handle} |
| Finance | {CFO/Founder} | @{handle} |
| Customer | {CS Lead name} | @{handle} |
## Meeting Types
- 1:1 — template: agenda / blockers / career
- standup — template: yesterday/today/blockers
- client — template: agenda / decisions / action items / sentiment
- sales — template: qualification / objections / next steps
- board — template: KPI review / strategic decisions / asks
## Default Owners
- Fallback: meeting organizer
- Reassignment window: 24h
Step 5l — competitors.md extension verification
Quick housekeeping step — no new questions for you. The agent re-opens the
competitors.mdfile generated earlier and verifies each competitor entry includes the new v2.4.0 fields (monitor_urls,track_pages,exec_team) thatcompetitor-monitorreads to do its weekly diff. If any competitor entry is missing these fields, the agent appendsTBDplaceholders. Thecompetitor-monitorskill will surface unfilled placeholders on its first run for you to complete then.Expect ~30 seconds. Just an automated verification.
Re-open the brands/{brand}/competitors.md file generated earlier in Step 5 and verify each competitor entry includes the monitor_urls, track_pages, and exec_team fields documented in the competitors.md template above. If any competitor entry is missing these fields (e.g. research-strategy output predated v2.4.0), append placeholders now:
- **monitor_urls:**
- homepage: {URL from existing entry}
- pricing: TBD
- blog: TBD
- careers: TBD
- changelog: TBD
- **track_pages:**
- pricing: price changes, new tier introductions
- homepage: hero copy / positioning shifts
- blog: new launches, thought-leadership themes
- careers: new exec hires
- changelog: new feature releases
- **exec_team:**
- TBD — fill manually from LinkedIn
The competitor-monitor skill will surface unfilled TBD placeholders on its first run so the user can complete them then; brand-setup does not block on this.
Step 6 — Logo
Quick step. The agent copies your logo file into brands/{brand}/logo.png so the visual skills (creative-designer, content-generator) can composite it onto generated images via Pillow. PNG with transparent background works best, but any image format the file system can read will do.
Expect 1 minute. Have your logo file ready at a known path. If you don't have a logo yet, you can skip this and add it later — text-only outputs work fine without a logo, and re-running brand-setup later to add the logo is non-destructive.
Make sure the logo file is somewhere I can read it. In Cowork that means inside your project folder (anywhere — root, a subfolder, doesn't matter, as long as it's inside the project mount). On local Claude Code, any path under your $HOME works.
Ask the user:
What is the file path to your logo? (PNG, transparent background preferred. Examples:
./logo.pngor./assets/logo.pngif you dropped it into your project — for Cowork, this needs to be inside your project mount.)
Copy the file from the provided path to brands/{brand}/logo.png:
import shutil
from pathlib import Path
raw = user_input.strip().strip('"').strip("'")
src = Path(raw).expanduser().resolve()
assert src.exists(), f"Logo not found: {src}"
assert src.is_file(), f"Not a file: {src}"
dst = Path("brands") / brand / "logo.png"
dst.parent.mkdir(parents=True, exist_ok=True)
shutil.copy2(src, dst)
If the path is invalid, tell the user what was wrong and ask again. This file is read by Python Pillow in content-generator and creative-designer for logo compositing.
Note: Google Font and brand colors were already discovered and saved to brands/{brand}/brand.md in Step 4.
Do not proceed to Step 7 until the logo file is confirmed copied (or the user explicitly skips it).
Step 7 — API Keys & Connections
Now we connect the integrations. This is where you'll spend most of the wall-clock time of brand-setup, because each integration takes a minute or two to authorize in Claude's Connectors UI or paste an API key. Most are click-through OAuth (Notion, Slack, Gmail, Google Calendar, Apollo, Stripe, Xero, Calendly, Zernio, etc.); a few require pasting an API key (FiveAgents gateway, Gemini) or entering credentials into a custom connector (DataforSEO — Basic Auth). Zernio is a click-through OAuth custom connector — no key to paste.
Expect 15–30 minutes depending on how many integrations you connect. Some are required (FiveAgents gateway, Notion, Slack), most are optional. The agent will tell you which is which as we walk through them, and you can say "skip" to any optional one — it'll just mark that skill as unconfigurable for this brand. You can always re-run brand-setup later to add a missing integration.
Walk through each integration one by one. For each one, explain what it does and whether it's required or optional. Ask: "Do you have your {integration} ready?" If the user says "not now" or "skip", acknowledge and move on — note it as unconfigured for the summary.
7a. Five Agents custom connector (MUST be first — all gateway tools depend on this):
Ask the user to add the Five Agents connector in Claude:
- Go to Settings → Connectors → "Add custom connector"
- Name:
Five Agents - URL:
https://gateway.fiveagents.io/api/mcp - Click Connect
This connector is required for all skills — it routes Gemini, email, and logging calls through the gateway. (Zernio and DataforSEO are no longer gateway-routed — as of gateway v1.7.4/v1.7.5 they each ship their own MCP server, connected separately in Steps 7a-ii and 7a-iii below.)
7a-ii. Zernio custom connector (required for social publishing):
Zernio (formerly Late) now runs its own hosted MCP server — it is no longer routed through the Five Agents gateway. Ask the user to add it in Claude:
- Go to Settings → Connectors → "Add custom connector"
- Name:
Zernio - URL:
https://mcp.zernio.com/mcp - Click Connect → sign in with your Zernio account (OAuth). No API key to paste.
If the user doesn't have a Zernio account yet, have them sign up at https://zernio.com first, then authorize. This must be done before content-generator or social-publisher can post.
7a-iii. DataforSEO custom connector (optional — keyword research):
DataforSEO also ships its own MCP server (no longer gateway-routed). Connect it only if the brand does keyword/SEO research (used by research-strategy and trend-radar):
- Go to Settings → Connectors → "Add custom connector"
- Name:
DataforSEO - URL:
https://mcp.dataforseo.com/mcp - Click Connect → authenticate with your DataforSEO API login + password (Basic Auth) from the DataforSEO dashboard → API Access. These are the same
DATAFORSEO_LOGIN/DATAFORSEO_PASSWORDcredentials captured in Step 7b.
7b. API Keys:
Required:
| # | Key | What it does | How to get it |
|---|---|---|---|
| 1 | FIVEAGENTS_API_KEY |
Dashboard logging, credential vault, email sending | Go to fiveagents.io → Dashboard → API Keys |
| 2 | GEMINI_API_KEY |
Image generation + text overlay (Google Fonts) | https://aistudio.google.com/apikey — free tier: 10 images/min |
| 3 | SLACK_NOTIFY_USER |
Slack DM notifications after each skill run | Open Slack → click your profile → three dots → "Copy member ID" |
| 4 | REPORT_EMAIL |
Email address for daily/weekly marketing reports | Your work email |
Required for social publishing: Zernio is connected as its own MCP connector in Step 7a-ii above (OAuth — no API key to store here).
Zernio setup (social media publishing):
Walk the user through connecting their social platforms to Zernio. This must be done before the content-generator or social-publisher skills can post.
To publish content to your social media accounts, we need to set up Zernio. Here's what to do:
Step A — Connect the Zernio MCP: Make sure you've added the Zernio connector in Step 7a-ii (Settings → Connectors → Add custom connector →
https://mcp.zernio.com/mcp→ sign in with your Zernio account). If you don't have a Zernio account yet, sign up at https://zernio.com first, then authorize. No API key to paste — it's OAuth.Step B — Create a Profile: In the Zernio dashboard, create a Profile for your brand. A profile groups all your social accounts together.
- Go to Profiles in the Zernio dashboard
- Click Create Profile
- Name it your brand name (e.g. "NPC Office")
Step C — Connect your social platforms: For each platform you want to publish to, connect it via OAuth in the Zernio dashboard:
- In your Profile, click Connect Account
- Select the platform (Facebook, Instagram, LinkedIn, etc.)
- Authorize Zernio to post on your behalf
- Repeat for each platform
Supported platforms: Facebook, Instagram, LinkedIn, Twitter/X, TikTok, YouTube, Pinterest, Threads, and more.
Important for Instagram: Instagram requires a Facebook Business Page linked to your Instagram Professional account. If you see an authorization error, check that your Instagram is set to Professional (not Personal) and is connected to a Facebook Page.
Important for LinkedIn: Connect your personal LinkedIn profile first. To post to a LinkedIn Company Page, you must be an admin of that page.
Let me know once you've connected your platforms and I'll automatically detect everything.
Step D — Auto-discover profile and account IDs:
After the user confirms they've connected their platforms, use the Zernio MCP to discover the profile ID and account IDs automatically:
1. Use Zernio MCP tool `profiles_list`:
→ Returns list of profiles. Pick the profile matching the brand name.
→ Save the profile `_id`.
2. Use Zernio MCP tool `accounts_list`:
- profile_id: "<profile _id from step 1>"
→ Returns list of connected accounts with `_id`, `platform`, `username`.
Show the user what was found:
I found your Zernio profile and these connected accounts:
- Facebook: @{username} (ID: {_id})
- Instagram: @{username} (ID: {_id})
- LinkedIn: {displayName} (ID: {_id})
- TikTok (if connected): @{username} (ID: {_id})
- Twitter/X (if connected): @{username} (ID: {_id})
- Google Ads: {platform} (SocialAccount ID: {_id})
- Meta Ads: {platform} (SocialAccount ID: {_id})
- LinkedIn Ads (if connected): {platform} (SocialAccount ID: {_id})
The accounts_list response includes all connected SocialAccount objects. Look for entries where platform is "googleads" or "google" for Google Ads, and "metaads" or "facebook" for Meta Ads. The _id on each is the SocialAccount ID (the Zernio-internal object ID).
Google Ads needs two IDs. Zernio's ad_campaigns_get_ads_timeline / ad_campaigns_list_ad_campaigns calls take both account_id (Zernio SocialAccount _id) and ad_account_id (the actual Google Ads customer ID — a 10-digit number). Passing only the SocialAccount ID returns empty results. Resolve the customer ID after saving the SocialAccount ID:
3. Use Zernio MCP tool `ads_list_ad_accounts`:
- account_id: <Google Ads SocialAccount _id from step 2>
→ Returns ad accounts visible under that Google Ads login.
3a. If accounts are returned → extract the numeric customer ID from the first entry. Save as ${BRAND}_ZERNIO_GOOGLE_ADS_CID.
3b. If the response is empty OR a 429 rate-limit error is returned → ask the user:
> "What is your Google Ads customer ID? You can find it in the top-right corner
> of your Google Ads account (format: XXX-XXX-XXXX)."
Strip dashes before saving — store only the digits.
LinkedIn Ads needs two IDs (same pattern as Google Ads). Per the Zernio LinkedIn Ads docs, ad_campaigns_get_ads_timeline / ad_campaigns_list_ad_campaigns for LinkedIn require both account_id (Zernio SocialAccount _id for the LinkedIn Ads entry) and ad_account_id (LinkedIn's numeric sponsored account ID, e.g. 517258773). The LinkedIn Ads SocialAccount is a separate accounts_list entry from organic LinkedIn — do not reuse {BRAND}_ZERNIO_LI (that is the organic LinkedIn page/profile ID for publishing).
4. From the same `accounts_list` response, find the LinkedIn Ads entry — `platform` is
"linkedinads" or "linkedin_ads" (distinct from the organic "linkedin" entry saved as
{BRAND}_ZERNIO_LI). Save its _id as ${BRAND}_ZERNIO_LINKEDIN_ADS.
5. Use Zernio MCP tool `ads_list_ad_accounts`:
- account_id: <LinkedIn Ads SocialAccount _id from step 4>
→ Returns LinkedIn sponsored ad accounts visible under that LinkedIn login.
5a. If accounts are returned → extract the numeric sponsored account ID (the `urn:li:sponsoredAccount:<id>`
digits) from the first entry. Save as ${BRAND}_ZERNIO_LINKEDIN_ADS_CID.
5b. If the response is empty OR a 429 rate-limit error is returned → ask the user:
> "What is your LinkedIn sponsored account ID? Open Campaign Manager → look at the URL or
> top-left account picker — it's a numeric ID (e.g. 517258773)."
Save the digits only.
Skip Steps 4–5 entirely if no LinkedIn Ads entry is found in accounts_list — most brands do not run LinkedIn Ads, and the analyst skills will silently skip LinkedIn analysis when these vars are unset.
Meta Ads only needs the single SocialAccount ID ({BRAND}_ZERNIO_META_ADS_ACCOUNT_ID) — its Zernio calls already work with just that.
Save the account IDs as env vars in .claude/settings.local.json (the content-generator, social-publisher, and video-repurposer skills need the social IDs; digital-marketing-analyst and data-analysis need the ads IDs for the Windsor.ai fallback):
{BRAND}_ZERNIO_FB → Facebook _id from accounts_list (platform: facebook)
{BRAND}_ZERNIO_IG → Instagram _id from accounts_list (platform: instagram)
{BRAND}_ZERNIO_LI → LinkedIn _id from accounts_list (platform: linkedin) — organic publishing only
{BRAND}_ZERNIO_TT → TikTok _id from accounts_list (platform: tiktok) — organic publishing (video-repurposer, YouTube-First)
{BRAND}_ZERNIO_TW → Twitter/X _id from accounts_list (platform: twitter or x) — organic publishing (video-repurposer, YouTube-First)
{BRAND}_ZERNIO_GOOGLE_ADS → Google Ads SocialAccount _id from accounts_list (platform: googleads or google)
{BRAND}_ZERNIO_GOOGLE_ADS_CID → Google Ads customer ID (10-digit, no dashes) from ads_list_ad_accounts or user input
{BRAND}_ZERNIO_META_ADS_ACCOUNT_ID → Meta Ads _id from accounts_list (platform: metaads or facebook)
{BRAND}_ZERNIO_LINKEDIN_ADS → LinkedIn Ads SocialAccount _id from accounts_list (platform: linkedinads or linkedin_ads) — distinct from {BRAND}_ZERNIO_LI
{BRAND}_ZERNIO_LINKEDIN_ADS_CID → LinkedIn sponsored account ID (numeric) from ads_list_ad_accounts or user input
Example: NPCOFFICE_ZERNIO_FB, NPCOFFICE_ZERNIO_IG, NPCOFFICE_ZERNIO_LI, NPCOFFICE_ZERNIO_TT, NPCOFFICE_ZERNIO_TW, NPCOFFICE_ZERNIO_GOOGLE_ADS, NPCOFFICE_ZERNIO_GOOGLE_ADS_CID, NPCOFFICE_ZERNIO_META_ADS_ACCOUNT_ID, NPCOFFICE_ZERNIO_LINKEDIN_ADS, NPCOFFICE_ZERNIO_LINKEDIN_ADS_CID
Only create env vars for platforms that were found. If a platform isn't connected to Zernio, skip that env var — the skill will note the gap if Windsor.ai fallback is triggered. For Google Ads and LinkedIn Ads, both vars (_ZERNIO_GOOGLE_ADS + _ZERNIO_GOOGLE_ADS_CID, _ZERNIO_LINKEDIN_ADS + _ZERNIO_LINKEDIN_ADS_CID) must be set for the fallback to work — if the customer/sponsored-account ID can't be obtained, save the SocialAccount ID alone and note the gap.
Save the profile ID and connected platforms to brands/{brand}/brand.md:
## Social Publishing
- Zernio Profile ID: {profile_id}
- Connected platforms: {Facebook (@username), Instagram (@username), LinkedIn (name), etc.}
Optional (ask for each — skip if not ready):
| # | Key | What it does | How to get it |
|---|---|---|---|
| 5 | DATAFORSEO_LOGIN |
Keyword research & search volume | https://dataforseo.com — sign up, copy login email |
| 6 | DATAFORSEO_PASSWORD |
Keyword research & search volume | DataforSEO dashboard → API Settings → API password |
| 7 | FIVEBUCKS_API_KEY |
The brand's fb.ai stack — paid fb.ai subscription (see Step 4b). Social templates (Step 4c) + SEO research, article generation, publishing, site audits, traffic monitoring, cold-email lead gen | https://www.fivebucks.ai/dashboard/api-keys — sign in (profile menu → API Keys), leave all 10 capability checkboxes ticked, generate, save it somewhere safe. See the scope note below. |
| 8 | {BRAND}_N8N_PROJECT |
Targets a specific n8n project/folder for n8n-workflow-builder's demo workflows |
n8n Cloud → open the target project → copy its project ID/name. Skip → workflows land in the n8n account's default project. Only relevant if you connected n8n Cloud. |
⚠️ fb.ai API key — the capability checkboxes
The fb.ai key page shows 10 capability checkboxes, and the key only grants what's ticked. They are all ticked by default — leave them that way unless the user deliberately wants to restrict what Claude can do.
Tell the user, in plain terms:
The key is scoped to one project and grants Claude the capabilities you tick. Ticking everything is the normal setup — it's what the skills expect. Only untick something if you specifically don't want Claude touching it.
| Checkbox | Unlocks | Needed by |
|---|---|---|
Social Posts (social_posts) |
Brand kit, templates, media, post drafts, render/export | Every skill that produces a visual — content-generator, creative-designer, social-publisher, social-calendar, and any skill that reads the brand kit (decks, reports). Untick it and ~12 skills stop working. |
| Integrations · SEO Research · Content · Publishing · Site Audit · Traffic Monitor | SEO research → article generation → publishing to 11 platforms; site audits; Search Console + AI-visibility tracking | The traffic skills: seo-researcher, article-publisher, site-auditor, traffic-reporter. These need no social_posts |
| Lead Gen Setup · Lead Gen CRM · Lead Gen Campaigns | Sending domains, lead search/enrichment, cold-email campaigns | The lead-gen skills |
Two things worth saying to the user up front:
- If a skill later fails with
auth/insufficient-scope, the key is missing that capability. Fix: regenerate the key at the same page with the box ticked, then re-run this step to re-store it. - The key can expire — the page offers an expiry and defaults to 90 days. If tools start failing with
auth/key-expired, generate a fresh key. Choose "No expiry" if the brand runs unattended automations and you don't want to revisit this.
Save ALL keys to .claude/settings.local.json:
For every key the user provides in Step 7b (including FIVEAGENTS_API_KEY, GEMINI_API_KEY, SLACK_NOTIFY_USER, REPORT_EMAIL, and any optional keys), save it to .claude/settings.local.json under the "env" object using the exact env var name shown in the tables above. This is required — all skills read credentials from env vars at runtime. (Zernio needs no key — it's an OAuth connector; DataforSEO's DATAFORSEO_LOGIN/DATAFORSEO_PASSWORD are captured here and also entered into the DataforSEO connector's Basic Auth.)
Also save these two vars — required for all automated skills:
DEFAULT_BRAND → the brand slug (e.g. "five-agents", "npc-office") — used by all skills to determine the active brand without user input
{BRAND}_NOTION_DB → the Notion Social Calendar DB page ID (a 32-character hex string from the DB's share URL) — required by social-calendar and content-generator
To get the Notion DB ID: open the Social Calendar database in Notion → click Share → Copy link → the 32-character hex string in the URL is the page ID. Save it as {BRAND}_NOTION_DB where {BRAND} is the brand slug uppercased with hyphens removed (e.g. FIVEAGENTS_NOTION_DB, NPCOFFICE_NOTION_DB).
Auto-bootstrapped Notion DB env vars (no user action at setup time):
Several skills each maintain their own Notion database for state — CRM, customer health, invoice tracker, financial reports, competitor monitor, meeting transcripts, action items (all v2.4.0), plus content-performance store and trend-radar candidates (v2.13.0). These DB env vars are auto-bootstrapped on first run of the relevant skill — the skill creates the database in the user's Notion workspace, captures the resulting page ID, and writes it to .claude/settings.local.json. No user action is required at setup time. Just acknowledge they exist; skills handle creation:
| Env Var | Bootstrapped By |
|---|---|
${BRAND}_CRM_DB |
apollo-lead-prospector / outreach-sequencer / proposal-generator (whichever runs first) |
${BRAND}_CUSTOMER_DB |
customer-onboarder / churn-predictor (whichever runs first) |
${BRAND}_INVOICE_TRACKER_DB |
invoice-collector |
${BRAND}_REPORTS_DB |
financial-reporter / investor-update-writer (whichever runs first) |
${BRAND}_COMPETITOR_DB |
competitor-monitor |
${BRAND}_MEETINGS_DB |
meeting-analyzer |
${BRAND}_ACTIONS_DB |
meeting-analyzer |
${BRAND}_PERFORMANCE_DB |
content-performance-analyst |
${BRAND}_TREND_DB |
trend-radar |
${BRAND}_GIGS_DB |
gig-prospector (the Inbound Gig Engine's shared opportunities DB; gig-proposal-writer / n8n-workflow-builder / vsl-demo-producer extend it in place) |
If the user later wants to point a skill at a pre-existing Notion DB instead of letting it auto-create one, they can paste the DB page ID into .claude/settings.local.json under the matching env var name before the skill's first run — the skill will detect the existing var and skip the bootstrap step.
Also store API keys in the credential vault (for Cowork use):
After saving ALL keys to settings.local.json, store each external API key in the encrypted vault so the gateway can access it:
For each key provided, use gateway MCP tool `fiveagents_store_credential`:
- fiveagents_api_key: ${FIVEAGENTS_API_KEY}
- service: "<service_name>"
- key: "<the_api_key>"
Use these service names (must match what the gateway expects):
| Env Var | Service Name |
|---|---|
GEMINI_API_KEY |
gemini |
FIVEBUCKS_API_KEY |
fivebucks |
Note: FIVEAGENTS_API_KEY, SLACK_NOTIFY_USER, and REPORT_EMAIL do NOT need vault storage — they are passed directly as tool parameters or used by built-in MCP connectors. They still MUST be saved to .claude/settings.local.json (done above).
Note: Zernio and DataforSEO are no longer gateway-vaulted (they left the gateway in v1.7.4/v1.7.5). Zernio authenticates via its own OAuth connector; DataforSEO authenticates via its own connector's Basic Auth using DATAFORSEO_LOGIN / DATAFORSEO_PASSWORD. Both are still saved to .claude/settings.local.json (DataforSEO's creds are entered into the connector; Zernio needs nothing stored).
Note: Google Ads, GA4, and Meta Ads (Facebook + Instagram) credentials are all handled by the Windsor.ai MCP connector — no gateway storage needed. The optional Meta Ads custom connector (https://mcp.facebook.com/ads) likewise authenticates via OAuth and stores nothing in the gateway vault. The Meta Ads MCP is in limited rollout — when available, downstream skills prefer it; when not, Windsor.ai already covers Meta data.
Keys are encrypted via Supabase Vault and can never be retrieved after storage. If the user needs to update a key later, they can re-run this step or use the dashboard UI at fiveagents.io.
7c. MCP Connectors:
Walk the user through each one. Explain what it does, ask the user to confirm they've connected it. If "not now", move on and note as unconfigured.
Meta Ads — Windsor.ai is the standard path; Meta Ads MCP is an optional enhancement:
Meta Ads (Facebook + Instagram) data is always pulled through Windsor.ai (source: "facebook") because Windsor is universally available and exposes near-parity with Meta's Marketing API — campaign / ad-set / ad / lp_views / video_views / conversions are all surfaced under their Windsor field names. Connecting Meta Ads inside the Windsor.ai dashboard is mandatory for every brand, and is part of the standard Windsor.ai Connected App setup further down this step.
Meta also ships an official MCP server at https://mcp.facebook.com/ads. It is currently in limited rollout — many accounts can't add it yet. Treat it as an optional enhancement: when the user has access, downstream skills will prefer it (it queries Meta's Marketing API directly with no Windsor middle-layer). When the user doesn't have access, Windsor.ai already covers everything those skills need — there is no degraded mode.
Optional — try the Meta Ads MCP (skip without prejudice if the user can't access it):
- In Cowork, go to Customize → Connectors → "Add custom connector"
- Name:
Meta Ads - URL:
https://mcp.facebook.com/ads - Click Connect and sign in with the Facebook/Meta Business account that owns the brand's ad accounts
Ask:
(Optional) Meta has an official MCP server you can also try, which gives skills slightly more direct access to the Marketing API. Want to give it a shot? If "Add custom connector" isn't visible, or sign-in fails with a "not available" error, no problem — your Windsor.ai connection (next) already covers Meta Ads fully.
- If the user adds it successfully → save
META_ADS_SOURCE=meta_ads_mcpto.claude/settings.local.jsonenvblock. Downstream skills (digital-marketing-analyst,data-analysis) will prefer the MCP for Meta data. - If the user skips or can't add it → leave
META_ADS_SOURCEunset. Downstream skills default to Windsor for Meta data, which is the universally-supported path.
⚠️ Connecting Meta Ads inside Windsor.ai is required regardless of whether the optional MCP was added — never skip it. The MCP, when present, is layered on top of (not in place of) the Windsor connection.
Connected Apps (OAuth via Settings → Connected Apps):
| # | MCP | What it does | How to connect |
|---|---|---|---|
| 1 | Notion | Content calendar, strategies & briefs | Settings → Connected Apps → Notion → Authorize |
| 2 | Slack | Notifications after each skill run | Settings → Connected Apps → Slack → Authorize |
| 3 | Gmail | Reading emails + report delivery | Settings → Connected Apps → Gmail → Authorize |
| 4 | Google Calendar | Scheduling content drops and meetings | Settings → Connected Apps → Google Calendar → Authorize |
| 5 | Windsor.ai (required) | Google Ads + GA4 + Meta Ads (Facebook + Instagram) — all three are mandatory. Connecting Meta Ads in Windsor is required regardless of whether the optional Meta Ads MCP was added above. | 1. Sign up for a free account at https://windsor.ai/register (if you don't have one yet)<br>2. In Windsor dashboard, connect Google Ads, GA4, and Meta Ads (Facebook Ads)<br>3. Then in Claude: Settings → Connected Apps → Windsor.ai → Authorize |
| 6 | Canva | Campaign presentations and pitch decks | Settings → Connected Apps → Canva → Authorize |
For Notion, Slack, Gmail, Google Calendar, and Canva, ask:
Have you connected {MCP name} in your Claude settings? (Settings → Connected Apps)
For Windsor.ai specifically, walk the user through all 3 steps before asking if they're done:
- "First, do you have a Windsor.ai account? If not, sign up free at https://windsor.ai/register"
- "Once you have an account, go to your Windsor dashboard and connect three accounts: Google Ads, GA4, and Meta Ads (Facebook Ads). All three are required — Meta Ads in Windsor is the universal source for Meta data, regardless of whether you also added the optional Meta Ads MCP earlier."
- "Then in Claude, go to Settings → Connected Apps → Windsor.ai → Authorize"
If yes to all, proceed. If "not now", acknowledge and move on.
Business-operations MCPs (added in v2.4.0 — required by the 10 new business-ops skills):
These MCPs are the connection layer for the v2.4.0 business-operations skills. You must walk the user through each one explicitly — do not just hand them the table and assume they'll connect what they need. Each prompt below explains what the MCP unlocks and what happens if they skip, so the user can make an informed call. Accept "skip" for any of them; that just marks the dependent skill(s) as unconfigurable for this brand.
| # | MCP | Used by | Skip behavior |
|---|---|---|---|
| 7 | Apollo.io | apollo-lead-prospector, outreach-sequencer |
Both skills become unconfigurable for this brand |
| 8 | Calendly | outreach-sequencer, customer-onboarder |
outreach-sequencer falls back to plain booking URL in sequence; customer-onboarder falls back to manual scheduling |
| 9 | Stripe | invoice-collector, financial-reporter |
Both skills become unconfigurable for this brand |
| 10 | Xero | invoice-collector (invoice sync), financial-reporter (P&L) |
financial-reporter falls back to Stripe-only data |
| 11 | PostHog (optional) | churn-predictor (product-usage signals) |
churn-predictor falls back to support-ticket + login-frequency signals only |
| 12 | Gamma (optional) | investor-update-writer (monthly investor decks) |
investor-update-writer becomes unconfigurable for this brand |
| 13 | n8n Cloud (optional) | n8n-workflow-builder (proof-of-concept workflow for inbound gig bids) |
n8n-workflow-builder becomes unconfigurable; the rest of the gig pipeline (gig-prospector → gig-proposal-writer) still runs |
Walk the user through each one in order. Post the prompt verbatim, wait for "done" or "skip", then move on.
Apollo.io:
Apollo.io powers lead enrichment, contact search, and email-sequence injection — used by
apollo-lead-prospectorandoutreach-sequencer. Connect: Settings → Connected Apps → Apollo.io → Authorize, then tell me when done. (Skip if you don't run outbound — those two skills will be unconfigurable for this brand.)Calendly:
Calendly handles booking links in outreach sequences and kickoff scheduling — used by
outreach-sequencerandcustomer-onboarder. Connect: Settings → Connected Apps → Calendly → Authorize, then tell me when done. (Skip if you don't use Calendly —outreach-sequencerfalls back to a plain booking URL andcustomer-onboarderfalls back to manual scheduling.)Stripe:
Stripe is the source of truth for invoice status and payment data — used by
invoice-collectorandfinancial-reporter. Connect: Settings → Connected Apps → Stripe → Authorize, then complete the OAuth flow it gates the real tools behind. Tell me when bothauthenticateandcomplete_authenticationhave succeeded. (Skip if you bill outside Stripe — those two skills will be unconfigurable for this brand.)Xero:
Xero is used for invoice status sync and pulling the P&L — used by
invoice-collectorandfinancial-reporter. Connect: Settings → Connected Apps → Xero → Authorize, then tell me when done. (Skip if you don't use Xero —financial-reporterfalls back to Stripe-only data.)PostHog (optional):
PostHog supplies product-usage signals (login frequency, feature engagement) for
churn-predictor. Connect: Settings → Connected Apps → PostHog → Authorize, then tell me when done. (Skip if you don't use PostHog —churn-predictorfalls back to support-ticket + login-frequency signals only.)Gamma (optional):
Gamma generates investor decks for
investor-update-writer's monthly cron run. Connect: Settings → Connected Apps → Gamma → Authorize, then tell me when done. (Skip if you don't deliver investor updates —investor-update-writerwill be unconfigurable for this brand.)n8n Cloud (optional):
n8n Cloud builds the proof-of-concept automation that backs an inbound gig bid — used by
n8n-workflow-builder(the Prove phase of the Inbound Gig Engine). Connect: Settings → Connected Apps → n8n Cloud → Authorize, then tell me when done. (Skip if you don't pursue freelance/marketplace work, or don't demo automations —n8n-workflow-builderwill be unconfigurable, butgig-prospectorandgig-proposal-writerstill run.)
Do not proceed to Step 8 until the user has responded to every integration in Step 7 — either configured or explicitly skipped. This includes the 7 business-ops MCPs above; if you skip them silently, Step 8c-bis will mark them ❌ and the user will hit "MCP not connected" errors on the first run of each dependent skill.
Step 8 — Validate Connections
Now we test every integration we just configured. The agent makes one cheap call per connector (e.g., whoami for Xero, list_labels for Gmail) and reports pass/fail. Anything that fails gets surfaced so you can fix it before relying on it in production — much better to find a bad credential now than discover it during a 6am cron run.
Expect ~5 minutes. No questions for you (mostly) — the agent runs the probes and shows a summary table. If any required integration fails, you'll be asked to re-authorize and the agent re-tests.
This step is mandatory and must not be skipped. Run validation for every integration the user configured in Step 7. For each test, show ✅ or ❌ with a clear error message if it fails. Only skip a specific test if the user explicitly chose not to set up that integration in Step 7.
8a. Gateway connector (MUST pass — all other gateway tests depend on this):
- Five Agents gateway — Log a test run:
Use gateway MCP tool `fiveagents_log_run`:
- fiveagents_api_key: ${FIVEAGENTS_API_KEY}
- skill: "brand-setup"
- brand: "{brand}"
- status: "success"
- summary: "Brand setup validation for {brand}"
- started_at: "<ISO timestamp>"
- completed_at: "<ISO timestamp>"
- metrics: { "date": "YYYY-MM-DD", "brand": "{brand}", "step": "validation" }
If error → tell user to verify their FIVEAGENTS_API_KEY and that the Five Agents connector URL is correct (https://gateway.fiveagents.io/api/mcp). Do NOT proceed with other gateway tests until this passes.
8b. Gateway API key tests (only if configured in Step 7):
- Credential vault — Store a test value and confirm it works:
Use gateway MCP tool `fiveagents_store_credential`:
- fiveagents_api_key: ${FIVEAGENTS_API_KEY}
- service: "test"
- key: "validation-check"
- Email sending — Send a test email:
Use gateway MCP tool `fiveagents_send_email`:
- fiveagents_api_key: ${FIVEAGENTS_API_KEY}
- to: ${REPORT_EMAIL}
- subject: "✅ Five Agents connected — {brand}"
- html_body: "<p>Your Five Agents plugin is set up for <strong>{brand}</strong>.</p>"
- tag: "brand-setup-test"
- Gemini (if
GEMINI_API_KEYconfigured):
Use gateway MCP tool `gemini_generate_text`:
- fiveagents_api_key: ${FIVEAGENTS_API_KEY}
- prompt: "Say hello"
- model: "gemini-2.5-flash"
- Image text overlay (if
GEMINI_API_KEYconfigured — tests Python Pillow):
from PIL import Image, ImageDraw
img = Image.new('RGB', (100, 100), color='gray')
draw = ImageDraw.Draw(img)
draw.text((10, 10), "Test", fill='white')
img.save('/tmp/test_overlay.png')
# If no error, Pillow is working
- Image logo overlay (if logo was provided in Step 6 — tests Python Pillow composite):
from PIL import Image
img = Image.new('RGBA', (100, 100), color='gray')
logo = Image.open('brands/{brand}/logo.png').convert('RGBA')
logo = logo.resize((20, 20), Image.LANCZOS)
img.paste(logo, (5, 5), logo)
img.save('/tmp/test_logo.png')
# If no error, logo compositing is working
- Zernio (if the Zernio connector is authorized):
Use Zernio MCP tool `profiles_list`:
→ Verify at least one profile exists
Use Zernio MCP tool `accounts_list`:
- profile_id: "<brand's profile ID>"
→ Verify connected accounts match what's in brands/{brand}/brand.md Social Publishing section
- DataforSEO (if the DataforSEO connector is authorized):
Use DataforSEO MCP tool `keywords_data_google_ads_search_volume`:
- keywords: ["test"]
- location_code: <infer from brand timezone/country — e.g. Indonesia=2360, Singapore=2702, Malaysia=2458, US=2840>
8c. MCP connectors (only if connected in Step 7):
- Slack (if connected) — Send a test DM:
slack_send_message to channel $SLACK_NOTIFY_USER:
"✅ Link plugin connected successfully for brand: {brand}"
Notion (if connected) — Try
notion-searchfor any page. If it returns results, Notion is connected.Gmail (if connected) — Try
gmail_get_profile. If it returns the user's email, Gmail is connected.Google Calendar (if connected) — Try
gcal_list_calendars. If it returns calendars, Google Calendar is connected.Windsor.ai (mandatory) — Try Windsor MCP
get_connectors. Verify the result includes all three required connectors: Google Ads, GA4, and Facebook (Meta Ads). All three are non-skippable — if any are missing, ask the user to connect the missing one in their Windsor dashboard before continuing. Mark ❌ if any of the three is absent; do not pass partial.Meta Ads MCP (only if
META_ADS_SOURCE=meta_ads_mcp— user opted into the optional MCP) — Call a basic listing tool on the Meta Ads connector (e.g. list ad accounts) and confirm it returns the user's Meta ad accounts without an auth error. If the call fails with an authorization error, ask the user to re-sign-in to the Meta Ads custom connector. IfMETA_ADS_SOURCEis unset (user did not add the optional MCP, or couldn't), mark this row as ⏭ skipped — Meta data is fully validated through the Windsor.ai check above and downstream skills will use Windsor for Meta automatically.Canva (if connected) — Try
list-brand-kits. If it returns results (even empty), Canva is connected.GA4 event discovery (if Windsor.ai connected AND funnel.md has TBD events) — Discover the client's actual GA4 conversion events:
Use Windsor.ai MCP tool `get_fields`:
- source: "googleanalytics4"
Look for key events and custom events (e.g. click_whatsapp, click_email, schedule_call, signup_form_submit, trial_activated, purchase, etc.). Filter out standard GA4 events (page_view, session_start, first_visit) — focus on key events (formerly called "conversions") and custom events that look like conversion actions.
Show the user the events found:
"I found these key events / custom events in your GA4 account: [list]. Which of these are your conversion actions for the funnel?"
After the user confirms, update brands/{brand}/funnel.md — replace any TBD event names with the confirmed GA4 event names.
8c-bis. Business-operations MCPs (only if connected in Step 7):
These probes validate the MCPs the business-operations skills depend on. Run each only if the user added the connector in Step 7. Mark ⏭ skipped if the user explicitly declined a given MCP (e.g. no Apollo account, no Stripe).
Apollo.io (used by
apollo-lead-prospector,outreach-sequencer) — Tryapollo_users_api_profile(cheapest call, just verifies OAuth is alive). If it returns the authenticated user, Apollo.io is connected. If it 401s, ask the user to re-authorize the Apollo connector in Settings → Connected Apps.Calendly (used by
outreach-sequencerandcustomer-onboarder) — Calendly's MCP exposes anauthenticateflow; once OAuth completes, call the basic profile / current-user lookup the connector exposes (e.g.users-get_current_userif available, otherwise rely on the OAuth-complete signal). If the user can't OAuth, they can paste a static Calendly link intobrands/{brand}/sales.mdReply Routing section instead — mark probe⏭ skippedand note in the email.Stripe (used by
invoice-collector,financial-reporter) — Stripe's MCP commonly only exposesauthenticate/complete_authenticationuntil OAuth completes. Check thatcomplete_authenticationsucceeded (or that an env-var-based Stripe key exists if the user opted into key auth). If neither, mark❌and ask the user to complete the OAuth flow before running invoice-collector.Xero (used by
invoice-collectorto sync invoice status,financial-reporterto pull P&L) — Try Xero MCPwhoami(orget_connected_user_organisation). If it returns the user's organisation, Xero is connected. If it 401s, ask the user to re-authorize.PostHog (used by
churn-predictorfor product-usage signals; data-analysis already validates PostHog if connected — reuse that result) — If not already validated, tryuser-get. If it returns the authenticated user, PostHog is connected. Mark⏭ skippedif the user has no PostHog account (churn-predictor will then fall back to support-ticket and login-frequency signals only).Gamma (used by
investor-update-writerfor investor decks; campaign-presenter already validates Gamma if connected — reuse that result) — If not already validated, tryget_themes. If it returns the user's themes, Gamma is connected.n8n Cloud (used by
n8n-workflow-builderfor the inbound-gig proof workflow; optional) — If connected, tryget_workflow_best_practices(technique="list") orget_sdk_reference. If it returns SDK/technique data, n8n Cloud is connected. Mark⏭ skippedif the brand doesn't pursue marketplace work —n8n-workflow-builderwill be unconfigurable, butgig-prospectorandgig-proposal-writerstill run.
8d. Workspace env vars (mandatory — required by automated skills):
DEFAULT_BRAND— Confirm.claude/settings.local.jsonenvblock containsDEFAULT_BRANDset to the brand slug (e.g."five-agents","npc-office"). If missing, ask the user for the brand slug and save it now. Required by every skill to determine the active brand without user input on scheduled runs.{BRAND}_NOTION_DB— Confirm.claude/settings.local.jsonenvblock contains{BRAND}_NOTION_DB(e.g.FIVEAGENTS_NOTION_DB,NPCOFFICE_NOTION_DB—{BRAND}is the slug uppercased, hyphens removed) set to the 32-character hex page ID of the brand's Notion Social Calendar database. If missing, walk the user through it: Notion → open the Social Calendar database → click Share → Copy link → extract the 32-hex-char ID from the URL → save it now. Required by social-calendar and content-generator.
Both env var checks are mandatory — they are NOT skippable. If either is missing, do not show ⏭ in the summary table; show ❌ and stop until the user provides the value.
After all tests, show a summary table:
| Integration | Status |
|---|---|
| Five Agents gateway | ✅ / ❌ |
| Credential vault | ✅ / ❌ |
| Email (Postmark) | ✅ / ❌ |
| Gemini | ✅ / ❌ / ⏭ skipped |
| Image text overlay | ✅ / ❌ / ⏭ skipped |
| Image logo overlay | ✅ / ❌ / ⏭ skipped |
| Zernio | ✅ / ❌ / ⏭ skipped |
| DataforSEO | ✅ / ❌ / ⏭ skipped |
| Slack | ✅ / ❌ / ⏭ skipped |
| Notion | ✅ / ❌ / ⏭ skipped |
| Gmail | ✅ / ❌ / ⏭ skipped |
| Google Calendar | ✅ / ❌ / ⏭ skipped |
| Windsor.ai (Google Ads + GA4 + Meta Ads — all required) | ✅ / ❌ |
| Meta Ads MCP (optional enhancement) | ✅ / ❌ / ⏭ skipped |
| Canva | ✅ / ❌ / ⏭ skipped |
| Apollo.io (apollo-lead-prospector / outreach-sequencer) | ✅ / ❌ / ⏭ skipped |
| Calendly (outreach-sequencer / customer-onboarder) | ✅ / ❌ / ⏭ skipped |
| Stripe (invoice-collector / financial-reporter) | ✅ / ❌ / ⏭ skipped |
| Xero (invoice-collector / financial-reporter) | ✅ / ❌ / ⏭ skipped |
| PostHog (churn-predictor) | ✅ / ❌ / ⏭ skipped |
| Gamma (investor-update-writer) | ✅ / ❌ / ⏭ skipped |
| n8n Cloud (n8n-workflow-builder) | ✅ / ❌ / ⏭ skipped |
DEFAULT_BRAND env var |
✅ / ❌ |
{BRAND}_NOTION_DB env var |
✅ / ❌ |
Show the table to the user. If any tests failed, offer to retry or troubleshoot before moving on. Save all results — they are used in Step 8d to build the agent readiness matrix and in the Step 10 completion email after CLAUDE.md is wired in Step 9.
Do not proceed to Step 8d until every configured integration has been tested and the summary table has been shown to the user.
Step 8d — Agent Readiness Summary (business-friendly)
The integrations table you just saw answers a technical question: which connections passed / failed / were skipped? It's accurate but it's not what the brand owner actually wants to know. The owner wants the business answer: which agents will actually run on my brand starting today, which need a fix, and which did I skip? Step 8d turns the integrations table into that answer.
Expect 30 seconds. No questions — the agent computes the matrix from Step 8 probe results + brand-file presence + env-var presence, then prints it to chat.
This step is mandatory. It runs before the Step 10 email so the user sees the readiness picture during the chat session, not just in the email afterward.
8d-i. Translation table (technical → business labels)
Use these display names everywhere in the matrix and the email — never raw MCP names, env vars, or .md filenames in user-facing output.
| Technical name | Business label to display |
|---|---|
| Apollo.io MCP | Apollo (lead database) |
| Calendly MCP | Calendly (meeting scheduler) |
| Stripe MCP | Stripe (payments) |
| PayPal MCP | PayPal (revenue) |
| Xero MCP | Xero (accounting) |
| PostHog MCP | PostHog (product analytics) |
| Gamma MCP | Gamma (presentations) |
| Canva MCP | Canva (presentations) |
| Notion MCP | Notion (your workspace) |
| Slack MCP | Slack (your team chat) |
| Gmail MCP | Gmail |
| Google Calendar MCP | Google Calendar |
| Google Drive MCP | Google Drive |
| Windsor.ai MCP | Windsor (Google Ads + GA4 + Meta Ads) |
| Meta Ads MCP | Meta Ads (direct API) |
| Zernio MCP | Social publisher |
| DataforSEO MCP | Keyword research |
| gateway: Gemini | Image generator |
| gateway: fivebucks | Branded social templates |
| gateway: email | Email reports |
sales.md |
Sales playbook |
customer-success.md |
Customer Success playbook |
finance.md |
Finance playbook |
investors.md |
Investor playbook |
operations.md |
Meeting routing rules |
brand.md |
Brand profile |
audience.md |
Customer personas |
product.md |
Product & pricing |
competitors.md |
Competitor list |
funnel.md |
Conversion funnel |
${BRAND}_NOTION_DB |
Social Calendar database |
${BRAND}_CRM_DB |
CRM database |
${BRAND}_CUSTOMER_DB |
Customer database |
${BRAND}_INVOICE_TRACKER_DB |
Invoice tracker |
${BRAND}_REPORTS_DB |
Reports archive |
${BRAND}_COMPETITOR_DB |
Competitor tracker |
${BRAND}_MEETINGS_DB |
Meeting archive |
${BRAND}_ACTIONS_DB |
Action items list |
${BRAND}_PERFORMANCE_DB |
Content performance store |
${BRAND}_TREND_DB |
Trend radar candidates |
${BRAND}_ZERNIO_FB |
Connected Facebook account |
${BRAND}_ZERNIO_IG |
Connected Instagram account |
${BRAND}_ZERNIO_LI |
Connected LinkedIn account |
8d-ii. Status rules
For each of the 20 business agents (every skill in agents/link.md Skills table except brand-setup and plugin-update, which are setup skills not business agents), assign exactly one status:
| Status | When to assign | What it means to the user |
|---|---|---|
| ✅ Ready | Every required dep is connected / present (per the agent's Deps row in agents/link.md). Optional deps may be missing. |
Configured and available to run end-to-end |
| ⚠️ Works with limitations | Every required dep present, but a known optional dep that materially affects output is missing. The canonical case: financial-reporter ready but Stripe not connected → still runs, but MRR/ARR cannot be computed. |
Will run, but with reduced output. State the specific limitation in the row. |
| ❌ Not ready yet | At least one required dep is missing. | Will fail on first run. State the specific missing dep + the exact fix command. |
| ⏭ You skipped | A required brand context file was explicitly skipped during Step 5 (e.g. investors.md skipped because the brand has not raised funding; operations.md skipped because the user does not process meeting transcripts). The skill is therefore intentionally off for this brand. |
No action needed. |
Important: Distinguish ❌ from ⏭ carefully. If customer-success.md is missing because the user forgot to fill it in → ❌ (with action: "run /link-skills:brand-setup Step 5h"). If investors.md is missing because the user said no in Step 5j Step 0 consent gate → ⏭. The Step 5g–5k consent gates and Step 8c-bis "skipped" markers tell you which case applies.
Optional deps that trigger ⚠️ Degraded (not ❌):
financial-reporterwithout Stripe → "produces P&L from Xero, but cannot show MRR/ARR until you connect Stripe"digital-marketing-analystwithout Meta Ads MCP → no degradation flag (Windsor covers Meta fully; this is informational only, treat as ✅ Ready)meeting-analyzerwithoutoperations.md→ no degradation flag (action items just route to "Unassigned"; treat as ✅ Ready)creative-designer/content-generatorwithoutdesign-system/→ no degradation flag (brand.md fallback is fully functional; treat as ✅ Ready)
Anything else missing that's marked (opt) in the skill's manifest deps → no impact on status; it's truly optional.
8d-iii. Build and display the matrix
Read plugins/link-skills/skills-manifest.json (generated from each skill's SKILL.md frontmatter by scripts/gen_skills_index.py). For every skill except brand-setup and plugin-update, check its deps (the mcp / gateway / files / env arrays) against Step 8 probe results + Step 5/6/7 file/env presence, assign the status per 8d-ii rules, group by the skill's area, and translate every dep name through the 8d-i table.
Print the result to chat in this exact format:
Five Agents — Brand: {brand} — Readiness Summary
✅ READY ({N} of 20) — configured and available to run
Marketing: Content Generator · Social Publisher · Social Calendar ·
Background Generator · Creative Designer · Content Writer ·
Research & Strategy · Campaign Presenter · Data Analysis ·
Digital Marketing Analyst
Sales: Apollo Lead Prospector
Finance: Invoice Collector
Strategy: Competitor Monitor
⚠️ WORKS WITH LIMITATIONS ({N})
Financial Reporter — produces P&L from Xero, but cannot show MRR/ARR
until you connect Stripe
Fix: Settings → Connected Apps → Stripe → Authorize
❌ NOT READY YET ({N}) — fix these to unlock:
Outreach Sequencer — needs Calendly (meeting scheduler)
Fix: Settings → Connected Apps → Calendly → Authorize
Proposal Generator — needs Stripe (payments) and Gamma (presentations)
Fix: Settings → Connected Apps → Stripe + Gamma → Authorize
Customer Onboarder — needs Calendly and Stripe
Fix: Settings → Connected Apps → Calendly + Stripe → Authorize
Churn Predictor — needs PostHog (product analytics) and Stripe
Fix: Settings → Connected Apps → PostHog + Stripe → Authorize
OR skip — only relevant if you use PostHog
⏭ YOU SKIPPED ({N}) — no action needed
Investor Update Writer — you don't have outside investors
Meeting Analyzer — you don't process meeting transcripts
→ {N_ready} ready · {N_degraded} works with limitations · {N_not_ready} need fixing · {N_skipped} skipped
Group the ✅ Ready section by Area (Marketing / Sales / Customer Success / Finance / Strategy / Productivity) using each skill's area from the manifest. Sub-bucket only if more than 3 skills land in the same area; otherwise list them inline.
For ⚠️ and ❌ rows, always include the specific fix command — never just "X is not connected". The fix command should be copy-pasteable (e.g. "Settings → Connected Apps → Stripe → Authorize").
For ⏭ rows, state why the user skipped (referencing the consent gate they answered "no" to in Step 5j or 5k). Do not state a fix; this is intentional opt-out, not a gap.
Save the computed matrix as a structured object — Step 10 reads it directly into the email payload's new agent_readiness[] block. Schema:
{
"agent_readiness": [
{
"name": "Apollo Lead Prospector",
"skill_id": "apollo-lead-prospector",
"area": "Sales",
"category": "Sales",
"status": "ready | degraded | not_ready | skipped",
"status_label": "Ready | Works with limitations | Not ready yet | You skipped",
"description": "Daily prospect search → Notion CRM",
"connected_tools": ["Apollo (lead database)", "Notion (your workspace)"],
"missing": ["Apollo (lead database)"],
"fix": "Settings → Connected Apps → Apollo.io → Authorize",
"skip_reason": null
}
]
}
Use missing: [] and fix: null for ✅ Ready rows. Use skip_reason (string) instead of fix for ⏭ rows.
8d-iv. Derive connected_tools[] from the manifest
Do not maintain a separate mapping table — skills-manifest.json (generated from each skill's frontmatter) is the source of truth. For each skill, read its deps and:
- Take every entry in
deps.mcpanddeps.gateway(skipdeps.filesanddeps.env— those aren't user-facing tools). - Each entry is a tool name optionally followed by an annotation, e.g.
"PostHog (opt)"or"Stripe (opt — payment-link fallback …)". Read the leading tool name + whether it carries(opt), then translate the name through the 8d-i table to its business display name (e.g.Windsor.ai→Windsor (Google Ads + GA4 + Meta Ads);Gemini→Image generator). - Preserve
(opt)as a parenthetical hint in the displayed name — e.g.PostHog (opt)displays asPostHog (product analytics, optional). - For ✅ Ready rows, only list tools that actually passed validation in Step 8a–8c-bis. For ❌ Not ready or ⚠️ Degraded, list the full set — what's missing is conveyed by the
missing[]field. For ⏭ Skipped, list the full set for informational display.
This keeps the readiness matrix in lockstep with the skills: each skill's deps live in its own SKILL.md frontmatter, the manifest regenerates from it (scripts/gen_skills_index.py), and this step reads the manifest — no separate edit here. If the manifest is stale, plugin-update (and gen_skills_index.py --check) flag it.
Do not proceed to Step 9 until the readiness matrix has been printed to chat and saved as the structured object for Step 10.
Step 9 — Initialize Workspace CLAUDE.md
The Link agent only knows it should activate for this brand if your workspace's CLAUDE.md file embeds the agents/link.md content. This step writes (or updates) CLAUDE.md at your workspace root, embedding the full Link agent definition along with credential-loading boilerplate and your brand defaults. Once this is in place, every Claude Code session in this workspace auto-loads Link without you having to invoke anything.
Expect ~2 minutes. No questions — the agent locates agents/link.md, reads it, and patches your CLAUDE.md idempotently between BEGIN/END markers (so re-running brand-setup or upgrading link.md just refreshes the embedded content without breaking anything you've added manually).
This step is mandatory and must not be skipped. It ensures every future session in this workspace (including scheduled/automated runs) loads the Link agent identity and credentials automatically. It runs before the completion email in Step 10 so that any CLAUDE.md write failure is caught and surfaced in the email rather than silently leaving the workspace half-configured.
We embed the full content of agents/link.md directly into CLAUDE.md rather than referencing an absolute path. This way the workspace is self-contained — scheduled runs, fresh clones, and machines without the plugin installed all still get the agent identity, because Claude Code auto-loads CLAUDE.md at session start.
9a. Locate and read agents/link.md
The agent definition file is bundled with the plugin. Find it on disk and read its contents into a variable.
The skill runs inside the Cowork sandbox (Ubuntu Linux VM, regardless of host OS), so the canonical search location is $CLAUDE_CONFIG_DIR/**/agents/link.md. The host-OS patterns are kept as fallbacks for the rare case the skill is invoked outside Cowork (e.g. local Claude Code on the user's machine).
import glob, os
# Cowork-first: $CLAUDE_CONFIG_DIR is the canonical plugin root inside the sandbox
# (e.g. /sessions/<name>/mnt/.claude). $HOME/.claude is a secondary location.
config_dir = os.environ.get("CLAUDE_CONFIG_DIR")
patterns = []
if config_dir:
patterns.append(os.path.join(config_dir, "**/agents/link.md"))
patterns.extend([
os.path.expanduser("~/.claude/**/agents/link.md"), # Linux / Cowork sandbox $HOME
os.path.expandvars(r"%APPDATA%\Claude\**\agents\link.md"), # Windows host (Claude Code locally)
os.path.expanduser("~/Library/Application Support/Claude/**/agents/link.md"), # macOS host (Claude Code locally)
])
found = [f for p in patterns for f in glob.glob(p, recursive=True)]
if found:
link_md_path = os.path.abspath(os.path.realpath(found[0]))
link_md_content = open(link_md_path, encoding='utf-8').read()
else:
link_md_path = ""
link_md_content = ""
Why $CLAUDE_CONFIG_DIR first: Cowork runs all skill code inside a sandboxed Ubuntu VM where process.platform === "linux" always. The Windows %APPDATA% and macOS ~/Library/Application Support/Claude patterns will never match inside the sandbox (they're for the user's host OS, which isn't reachable from skill code). $CLAUDE_CONFIG_DIR is the env var Cowork sets to point at the mounted plugin tree (typically /sessions/<session>/mnt/.claude), so it's the only pattern guaranteed to find link.md inside Cowork. The ~/.claude pattern works in some Cowork configurations via bindfs mounts but isn't reliable.
If the search returns empty, ask the user:
I couldn't auto-detect
agents/link.md. Can you paste the full absolute path to it? (Hint: in Cowork, runecho $CLAUDE_CONFIG_DIRin a terminal and look under that directory; on local Claude Code, search your Claude application data folder — Windows%APPDATA%\Claude, macOS~/Library/Application Support/Claude, Linux~/.claude.)
After the user pastes a path, normalize, validate, and read:
user_path = os.path.abspath(os.path.expanduser(os.path.expandvars(user_input.strip().strip('"').strip("'"))))
assert os.path.isfile(user_path), f"File not found: {user_path}"
link_md_path = user_path
link_md_content = open(link_md_path, encoding='utf-8').read()
Strip the YAML frontmatter from link_md_content before embedding (the --- block at the top with name: / description:). The frontmatter is a plugin-loader directive and has no meaning inside CLAUDE.md:
import re
link_md_body = re.sub(r'^---\s*\n.*?\n---\s*\n', '', link_md_content, count=1, flags=re.DOTALL).lstrip()
Extract the version from the Maintenance table so it can be stamped into CLAUDE.md:
import datetime
version_match = re.search(r'\|\s*Link\s*\|\s*(v[\S]+)\s*\|\s*([^|\n]+)\s*\|', link_md_content)
link_version = version_match.group(1).strip() if version_match else "unknown"
link_version_date = version_match.group(2).strip() if version_match else "unknown"
embed_date = datetime.date.today().isoformat() # e.g. 2026-05-06
link_md_body is what gets embedded in 9b.
9b. Read or create CLAUDE.md
Check if CLAUDE.md exists at the workspace root (same folder as brands/ and outputs/).
Build the workspace block to inject. Substitute these placeholders verbatim:
{LINK_MD_BODY}— the stripped contents oflink_md_bodyfrom 9a{brand}— the brand slug from Step 3 (same value saved asDEFAULT_BRAND, e.g.five-agents){BRAND}— the brand slug uppercased with hyphens removed, used as the env var prefix (e.g.FIVEAGENTS,NPCOFFICE){notion_db_id}— the 32-character Notion Social Calendar DB page ID collected in Step 7b{link_version}— the version string extracted from link.md's Maintenance table (e.g.v2.3.0){link_version_date}— the Last Changed date from link.md's Maintenance table (e.g.May 06, 2026){embed_date}— today's date in ISO format (e.g.2026-05-06), set bydatetime.date.today().isoformat()
# {Brand Name} — Workspace Instructions
## Agent Identity (auto-loaded every session)
The full content of `agents/link.md` is embedded below. It defines your identity (Link), active brand logic, available skills, tools, integrations, output conventions, and quality checklist. All skill runs depend on it.
<!-- BEGIN agents/link.md (embedded by brand-setup) -->
<!-- link.md version: {link_version} | Last Changed: {link_version_date} | Embedded: {embed_date} -->
{LINK_MD_BODY}
<!-- END agents/link.md -->
---
## Credential Loading (REQUIRED — read this first on every run)
Scheduled and automated runs do **not** automatically inject environment variables from `.claude/settings.local.json`. You must load them manually at the start of every skill run using this snippet:
import os, json
from pathlib import Path
def load_credentials():
search = Path(os.getcwd())
for p in [search] + list(search.parents):
settings_file = p / ".claude" / "settings.local.json"
if settings_file.exists():
data = json.loads(settings_file.read_text())
for k, v in data.get("env", {}).items():
if not os.environ.get(k):
os.environ[k] = v
return True
return False
load_credentials()
Run this **before** reading any env var (`FIVEAGENTS_API_KEY`, `SLACK_NOTIFY_USER`, `{BRAND}_ZERNIO_FB`, etc.). If `FIVEAGENTS_API_KEY` is still missing after this step, log a `failed` run and exit — do not skip publishing silently.
---
## Workspace Defaults
These values are hardcoded here at brand-setup time so any session reading `CLAUDE.md` has them immediately — no env lookup required for the common path. They are also saved in `.claude/settings.local.json` `env` block as a fallback.
- **Active brand:** `{brand}` (env: `DEFAULT_BRAND`)
- **Brand files:** `brands/{brand}/`
- **Notion Social Calendar DB:** `{notion_db_id}` (env: `{BRAND}_NOTION_DB`)
**Fallback rule:** if either tagged value above looks empty or stale (e.g. a literal `{brand}` placeholder that never got substituted, or this `CLAUDE.md` was copied from another workspace), run the credential loader and read from env:
brand = os.environ["DEFAULT_BRAND"]
notion_db = os.environ[f"{brand.replace('-', '').upper()}_NOTION_DB"]
## Workspace Structure
brands/{brand}/ — brand context root
├─ brand.md, audience.md, product.md, — written by brand-setup Steps 4–5
│ competitors.md, funnel.md, logo.png
├─ sales.md — outbound sales config (v2.4.0 Step 5g — apollo-lead-prospector, outreach-sequencer, proposal-generator)
├─ customer-success.md — onboarding + retention config (v2.4.0 Step 5h — customer-onboarder, churn-predictor)
├─ finance.md — billing + financial-reporting config (v2.4.0 Step 5i — invoice-collector, financial-reporter)
├─ investors.md — investor-update config (v2.4.0 Step 5j — investor-update-writer; absent if no outside funding raised)
├─ operations.md — meeting-processing config (v2.4.0 Step 5k — meeting-analyzer; absent if not used)
├─ backgrounds/ — pre-generated background images (background-generator skill)
└─ design-system/ — Claude Design export (Step 4b — see Visual System below)
(social templates live on fb.ai, not here — Step 4c)
outputs/{brand}/ — all generated content (copy .md, images .png, videos .mp4)
tmp/ — scratch space for scripts, intermediate files
## Account IDs (Zernio)
Read from env vars after credential loading:
- Facebook: `{BRAND}_ZERNIO_FB`
- Instagram: `{BRAND}_ZERNIO_IG`
- LinkedIn (organic publishing): `{BRAND}_ZERNIO_LI`
- Google Ads Zernio account (Windsor.ai fallback): `{BRAND}_ZERNIO_GOOGLE_ADS`
- Google Ads customer ID (Windsor.ai fallback): `{BRAND}_ZERNIO_GOOGLE_ADS_CID`
- Meta Ads account (Windsor.ai fallback): `{BRAND}_ZERNIO_META_ADS_ACCOUNT_ID`
- LinkedIn Ads Zernio account (Windsor.ai fallback): `{BRAND}_ZERNIO_LINKEDIN_ADS`
- LinkedIn sponsored account ID (Windsor.ai fallback):`{BRAND}_ZERNIO_LINKEDIN_ADS_CID`
---
If CLAUDE.md already exists:
- If it contains the markers
<!-- BEGIN agents/link.md (embedded by brand-setup) -->and<!-- END agents/link.md -->, replace everything between (and including) those markers with the freshly read{LINK_MD_BODY}wrapped in the same markers and a refreshed version stamp comment (<!-- link.md version: {link_version} | Last Changed: {link_version_date} | Embedded: {embed_date} -->). Leave the rest of the file untouched. - If it contains an older
## Agent Identitysection that points to an absolute path (the previous format), replace the entire block from## Agent Identitydown through the---separator with the new workspace block above. - Otherwise, prepend the new workspace block above all existing content.
- Refresh the
## Workspace Defaultssection: rewrite Active brand, Brand files, and Notion Social Calendar DB to point at the brand from this run. For multi-brand workspaces, do NOT overwrite — append a new sub-block under## Workspace Defaultstitled### Brand: {brand}with the same three bullets, leaving prior brands' sub-blocks intact. TheDEFAULT_BRANDenv var still selects which brand is active per session.
If CLAUDE.md does not exist:
- Create it with the full workspace block above.
Show the user what was written:
✅
CLAUDE.mdupdated — the Link agent definition is now embedded directly in this workspace, so every future session loads it automatically (no plugin path lookups needed).
9c. Detect and wire brand visual assets (best-effort — non-mandatory)
After CLAUDE.md is written by 9b, detect the brand's visual assets. Design system is a local folder (brands/{brand}/design-system/, Step 4b) — the crucial, free baseline that always works. It can optionally also be uploaded to fb.ai (detect via fivebucks_get_brand_kit, needs FIVEBUCKS_API_KEY + paid plan). Social templates (Step 4c) and the media library (Step 4d) live on fb.ai only — detect via fivebucks_list_templates and fivebucks_list_media_folders (both need FIVEBUCKS_API_KEY).
| Asset | Where | Used by |
|---|---|---|
brands/{brand}/design-system/ |
local folder (Step 4b) — optionally mirrored to fb.ai | every visual-producing skill — colors, typography, components, spacing |
fb.ai social templates (meta-carousel / meta-story / linkedin-post / meta-post) |
fb.ai (Step 4c, via fivebucks_list_templates) |
content-generator + creative-designer render via fivebucks_*; content-creation reads the manifest to size copy |
| fb.ai media library | fb.ai (Step 4d, via fivebucks_list_media_folders) |
content-generator + creative-designer pull on-brand photos at runtime |
⚠️ Non-mandatory. All are optional (Steps 4b/4c/4d are recommended-but-skippable). This step never fails or blocks — it just records what's present so the email and CLAUDE.md reflect reality. Skills fall back to brand.md colors/fonts + Gemini + Pillow when an asset is absent. Never error out, never block the Step 10 completion email.
from pathlib import Path
brand_root = Path("brands") / brand # `brand` is the slug from Step 3
def folder_status(name):
p = brand_root / name
return "installed" if (p.is_dir() and any(p.iterdir())) else "not installed"
design_system_status = folder_status("design-system") # local — the crucial baseline
# Optional fb.ai mirror — call fivebucks_get_brand_kit; non-null means the design system
# was also uploaded to fb.ai (so fb.ai templates can render with brand colors).
# design_system_fbai = "uploaded" if result is non-null, else "not uploaded" (skip if no FIVEBUCKS_API_KEY)
# Social templates live on fb.ai — detect via the gateway (best-effort; skip if FIVEBUCKS_API_KEY unset).
# Call fivebucks_list_templates and collect each returned template's `type`.
fb_template_types = [] # e.g. ["meta-carousel", "linkedin-post"], or [] if none / no fb.ai key
# Media library — call fivebucks_list_media_folders; media_status = "installed" if any folders exist, else "not installed".
Build the Visual System block to inject into CLAUDE.md. Substitute {design_system_status}, {design_system_fbai}, {fb_template_types}, and {media_status}:
<!-- BEGIN visual-system (managed by brand-setup Step 9c) -->
## Visual System
Detected at brand-setup time. Re-run Step 9c (or full brand-setup) after installing/uploading more assets to refresh.
- **Design system:** `brands/{brand}/design-system/` — **{design_system_status}** (local source of truth for colors, typography, components, spacing — read by every visual-producing skill). fb.ai mirror: **{design_system_fbai}**.
- **Social templates (fb.ai):** {fb_template_types or "none"} — types present out of meta-carousel / meta-story / linkedin-post / meta-post
- **Media library (fb.ai):** **{media_status}** — brand photos available for post generation via `fivebucks_list_media_files`
**How skills use this:** the lookup order for brand colors/fonts is **fb.ai brand kit (`fivebucks_get_brand_kit`, when uploaded) → local `design-system/` → `brand.md`**. Social templates and media live on fb.ai only (`fivebucks_list_templates` / `fivebucks_list_media_files`). When fb.ai assets are absent — or `FIVEBUCKS_API_KEY` isn't set — skills use the local `design-system/`, then `brand.md` + Gemini + Pillow. The fallback is fully functional; it just lacks brand-specific layout chrome. This section is a hint, not a contract — skills re-check at runtime.
<!-- END visual-system (managed by brand-setup Step 9c) -->
Inject into CLAUDE.md idempotently:
- If
CLAUDE.mdalready contains the markers<!-- BEGIN visual-system (managed by brand-setup Step 9c) -->and<!-- END visual-system (managed by brand-setup Step 9c) -->, replace everything between (and including) those markers with the freshly built block. Leave the rest of the file untouched. - Otherwise, append the new block to the end of
CLAUDE.mdpreceded by a blank line.
Show the user a concise status line:
✅
CLAUDE.mdVisual System section refreshed:
- design-system (local): {design_system_status} · fb.ai mirror: {design_system_fbai}
- fb.ai social templates: {fb_template_types or "none"}
- media library: {media_status}
If design-system is not installed and no fb.ai templates are present, gently prompt:
Heads up — no brand visual assets are set up yet. Skills still work (they fall back to Gemini + Pillow), but for the most on-brand visuals, run Step 4b (design system), and optionally Step 4c (fb.ai social templates) + Step 4d (media library) when you have time.
This step is non-blocking and safe to re-run on its own at any later point — useful when the user finishes installing a template after initial brand-setup.
Do not proceed to Step 10 until 9b has written CLAUDE.md (9c is best-effort and may report not installed for any folder without blocking). If the 9b write itself failed, fix it first — Step 10 will report the actual write status in the email, but a hard write failure should be resolved before the user is told setup is complete.
Step 10 — Summary & Completion Email
Final step. The agent compiles a summary of everything that was done — which files were created, which integrations connected, which were skipped, which validations passed/failed — and ships it to two destinations: an HTML completion email to your $REPORT_EMAIL (via the gateway) and a Slack DM to $SLACK_NOTIFY_USER. Future-you (or whoever you handed this brand off to) gets a paper trail of what was set up and what's still pending.
Expect 1 minute. No questions. After this step finishes, you can run any of the 23 Link skills against this brand. The first scheduled cron run (typically social-calendar Sunday or apollo-lead-prospector daily) will surface any remaining gaps.
This step is mandatory and must not be skipped. It is the final step of brand-setup. Always send the completion email and Slack notification once Steps 8 (validation) and 9 (CLAUDE.md) have completed, regardless of how many integrations were configured.
Send a completion email to $REPORT_EMAIL with the full setup report.
⚠️ Do NOT generate HTML. Build a JSON object. The server-side template (brand-setup.ts) renders the styled email with tables, status badges, and callouts.
Use gateway MCP tool `fiveagents_send_email`:
- fiveagents_api_key: ${FIVEAGENTS_API_KEY}
- to: ${REPORT_EMAIL}
- subject: "✅ Brand setup complete — {brand}"
- html_body: JSON.stringify(payload)
- tag: "brand-setup"
⚠️ tag must be exactly "brand-setup" — this routes to the server-side template.
Build the JSON payload from Step 8d's saved readiness matrix (the primary block — the user reads this first), Step 8 validation results, and Step 9 CLAUDE.md / visual-asset status. The server-side template (brand-setup.ts) renders agent_readiness[] as the headline section; connections[] and files[] are rendered as collapsed "Diagnostic detail" sections for debugging.
Read the brand display name from brands/{brand}/brand.md before building the payload. The display name is the first # heading in that file (e.g. # Five Agents → "Five Agents"). Use it for the top-level brand_name field — the template renders it in the email title. Do NOT pass the slug (five-agents) — that goes in the separate brand field.
import re
brand_md = open(f"brands/{brand}/brand.md", encoding="utf-8").read()
m = re.search(r'^#\s+(.+?)\s*$', brand_md, re.MULTILINE)
brand_name = m.group(1).strip() if m else brand # fall back to slug only if heading missing
{
"brand": "{brand}",
"brand_name": "{brand_name}",
"agent_readiness": [
{
"name": "Apollo Lead Prospector",
"skill_id": "apollo-lead-prospector",
"area": "Sales",
"category": "Sales",
"status": "ready | degraded | not_ready | skipped",
"status_label": "Ready | Works with limitations | Not ready yet | You skipped",
"description": "Daily prospect search → Notion CRM",
"connected_tools": ["Apollo (lead database)", "Notion (your workspace)"],
"missing": ["Apollo (lead database)"],
"fix": "Settings → Connected Apps → Apollo.io → Authorize",
"skip_reason": null
}
// ... one entry per business agent (20 total — every link.md Skills row except brand-setup and plugin-update)
],
"readiness_summary": {
"ready": 0,
"degraded": 0,
"not_ready": 0,
"skipped": 0,
"total": 20
},
"files": [
{ "file": "brands/{brand}/brand.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/product.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/audience.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/competitors.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/funnel.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/sales.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/customer-success.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/finance.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/investors.md", "status": "present | skipped | failed", "notes": "skipped if brand has not raised outside funding" },
{ "file": "brands/{brand}/operations.md", "status": "present | skipped | failed", "notes": "skipped if user does not run meetings through Claude" },
{ "file": "brands/{brand}/logo.png", "status": "present | missing | failed" },
{ "file": "CLAUDE.md", "status": "present | missing | failed" },
{ "file": "brands/{brand}/design-system/", "status": "present | missing | failed" }
],
"connections": [
{ "integration": "Five Agents gateway", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Credential vault", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Email (Postmark)", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Gemini", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Image text overlay", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Image logo overlay", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Zernio", "status": "pass | fail | skipped", "notes": "Connected: Facebook, Instagram, LinkedIn" },
{ "integration": "DataforSEO", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Slack", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Notion", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Gmail", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Google Calendar", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Windsor.ai", "status": "pass | fail", "notes": "Required: Google Ads + GA4 + Meta Ads (Facebook + Instagram) all connected — universal source for paid ads + analytics" },
{ "integration": "Meta Ads MCP", "status": "pass | fail | skipped", "notes": "Optional enhancement — Marketing API direct access. When connected (META_ADS_SOURCE=meta_ads_mcp), downstream skills prefer it over Windsor for Meta data. When skipped or unavailable, Windsor.ai already covers Meta Ads fully." },
{ "integration": "Canva", "status": "pass | fail | skipped", "notes": "" },
{ "integration": "Apollo.io", "status": "pass | fail | skipped", "notes": "Used by apollo-lead-prospector / outreach-sequencer. Skip → those skills can't run." },
{ "integration": "Calendly", "status": "pass | fail | skipped", "notes": "Used by customer-onboarder and outreach-sequencer. Skip → fall back to a static link in sales.md." },
{ "integration": "Stripe", "status": "pass | fail | skipped", "notes": "Used by invoice-collector / financial-reporter. Skip → those skills can't run." },
{ "integration": "Xero", "status": "pass | fail | skipped", "notes": "Used by invoice-collector / financial-reporter. Skip → those skills can't run." },
{ "integration": "PostHog", "status": "pass | fail | skipped", "notes": "Used by churn-predictor for product-usage signals. Skip → churn scoring falls back to support-ticket + login-frequency only." },
{ "integration": "Gamma", "status": "pass | fail | skipped", "notes": "Used by investor-update-writer for investor decks. Skip → updates render as plain markdown / email only." },
{ "integration": "n8n Cloud", "status": "pass | fail | skipped", "notes": "Used by n8n-workflow-builder to build the proof-of-concept automation backing an inbound gig bid. Skip → that skill is unconfigurable; gig-prospector and gig-proposal-writer still run." },
{ "integration": "DEFAULT_BRAND env var", "status": "pass | fail", "notes": "Active brand slug — required by every skill (mandatory, not skippable)" },
{ "integration": "{BRAND}_NOTION_DB env var", "status": "pass | fail", "notes": "Notion Social Calendar DB page ID — required by social-calendar and content-generator (mandatory, not skippable)" }
],
"action_items": [
{ "integration": "{name}", "message": "{what failed or was skipped and how to fix / which skill needs it}" }
]
}
files[] status enum — most rows use three values; two rows (investors.md, operations.md) accept a fourth skipped value:
present— file or folder exists with expected content (the happy path; covers "newly created", "already existed and updated", and "installed by user").missing— file/folder absent because the relevant step was skipped or the user didn't supply input (e.g.logo.pngwhen the user skipped Step 6,design-system/when the user skipped Step 4b).skipped— used only forinvestors.md(when the brand hasn't raised outside funding) andoperations.md(when the user doesn't run meetings through Claude). These are legitimate "not applicable" outcomes — not failures. Do not raise an action item for these.failed— the write/copy/folder-probe operation raised an error during this run (e.g. permission denied, disk full, Step 9b couldn't write CLAUDE.md). Afailedvalue onCLAUDE.mdMUST also produce anaction_itemsentry pointing the user to retry Step 9.
Where each files[] status comes from:
.mdrows andlogo.png— set during the file-generation steps (3, 4, 5, 6). Usepresentif written,missingif the user skipped the step,failedif the write raised.sales.md,customer-success.md,finance.md— set by Steps 5g / 5h / 5i. These are mandatory for the v2.4.0 sales/CS/finance skills, somissingshould only appear when the user explicitly declined those workflows; otherwisepresentorfailed.investors.md— set by Step 5j. Usepresentif generated,skippedif the brand hasn't raised outside funding (the user said no in 5j Step 0),failedif the write raised.operations.md— set by Step 5k. Usepresentif generated,skippedif the user doesn't run meetings through Claude,failedif the write raised.CLAUDE.md— set by Step 9b:presentif the file is on disk with the expected workspace block (whether newly created or refreshed in place),failedif the write raised. (missingshould never appear since Step 9 is mandatory.)- The
design-system/row — set by Step 9c'sfolder_status()result:presentwhen the local folder exists and is non-empty,missingwhen absent or empty. (Social templates and the media library are not infiles[]— they live on fb.ai; the Visual System block records their presence viafivebucks_list_templatesandfivebucks_list_media_folders. The optional fb.ai brand-kit mirror is also recorded there, viafivebucks_get_brand_kit.)
connections[] status enum — pass | fail | skipped. Use "pass" for ✅, "fail" for ❌, "skipped" for ⏭. Only include action_items entries for failures and skips that actually affect skill functionality. Specifically:
- ✅ Do add an action item for any
fail(these always need user action). - ✅ Do add an action item for
skippedkeys/MCPs whose absence breaks a skill (e.g. Zernio connector not authorized → social-publisher can't post;GEMINI_API_KEYskipped → no image generation). - ✅ Do add an action item if
CLAUDE.mdcame backfailedfrom Step 9b — scheduled / automated runs depend on it. - ❌ Do NOT add an action item for skipped Meta Ads MCP — Windsor.ai already covers Meta data fully when the MCP isn't available, so a skip here is a no-op for downstream skills, not a gap.
- ❌ Do NOT add an action item for genuinely optional integrations the user explicitly declined (e.g. Apollo skipped because the brand doesn't run outbound, Stripe/Xero skipped because the brand doesn't have those accounts yet).
- ❌ Do NOT add an action item for visual asset folders showing
missing— Step 9c already nudged the user inline; padding the email with these rows isn't useful. - ❌ Do NOT add an action item for
investors.mdoroperations.mdshowingskipped— those are legitimate "not applicable" outcomes (no fundraising; no meeting workflow), not gaps. Add an action item only if those files came backfailed.
Also print the same summary to the chat and send a Slack notification to $SLACK_NOTIFY_USER. The Slack message leads with the readiness counts from readiness_summary (what users care about), with the integration count as a secondary diagnostic line:
✅ Brand "{brand}" setup complete
Agent Readiness (20 total):
✅ {N_ready} ready · ⚠️ {N_degraded} works with limitations · ❌ {N_not_ready} need fixing · ⏭ {N_skipped} skipped
{If N_not_ready > 0:}
Top fixes to unlock more agents:
• {Agent name 1} — {one-line fix from agent_readiness[].fix}
• {Agent name 2} — {one-line fix}
(full list in the email)
CLAUDE.md: {present | failed}
Brand files: brands/{brand}/
{N} action items (see email for full details)
Cap the "top fixes" inline list at 3 — the rest live in the email. If N_not_ready == 0, omit the "Top fixes" block entirely; lead instead with "🎉 Every connected agent is configured and ready to run."