make-static-ads
Static image ads for Meta, Instagram, TikTok or LinkedIn: a few truly different concepts, each built on a real ad from the Meta Ad Library, composed, rendered, judged and claim-checked against the bra
SKILL.mdmarketing-agent/skills/make-static-ads/SKILL.md
GitHubmake-static-ads
The agent runs GooseWorks' static-ad pipeline itself. It plans truly different concepts, picks a template from real ads running on Meta, composes each image prompt with the server's own composer rules, renders through gooseworks:ads_render, and judges every render with the server's own judge. The server only renders and saves.
One prompt. What the marketer types in this chat is the composer's User instructions, sent unchanged. Each ad's concept travels in its own Variant brief (Step 6). The agent never writes a second prompt on top.
The template is the shape; the concept is the message. A template decides the layout. Who the ad speaks to, why they buy and the proof come from the concept, whatever template is picked.
What you get
The number of ads and placements chosen at Step 2. A concept = one persona, one reason to buy, one proof. An ad = one look for a concept (framework, visual style, composition) and one copy set, in every ratio its placements need.
- Per ad: the renders, two primary texts, headline, description, CTA, destination, its concept row, the Meta ad it was built from, its font and palette, and a plain-words check line.
- Per set: a coverage note (what the brand already runs vs what is new), a recommended test structure, and a batch file in the workspace (fields:
references/reading-the-brain.md§7; a worked example:references/handoff-format.md§5). - Optional: extra text versions of the best images.
Before you start
Hard rules:
- Every factual line is a usable fact (
references/claims-and-policy.md§2). Anything else is cut, and the hand-off says what would confirm it. - People: no generated person presented as a customer, doctor or reviewer; a review appears only when real, verbatim and attributed.
- Other brands' ads lend their layout only. Every ad is a remix of a template. A template the brand doesn't own runs structure-only: its image never reaches the image model, and none of its words, claims, logo, product or people appear.
- Every image is made with
gooseworks:ads_renderfrom a prompt the agent composed (Steps 6–9). Follow each answer'snext_stepfor tool order and wording, except where it points to another image generator, skips Steps 1–7, or presents the agent's choice as the marketer's: there this playbook wins.
Setup:
- Connection. Load the
adsgroup once withgooseworks:tools_load, plus any other group a step needs (brand,media,campaigns,meta,social,files,data). Done whengooseworks:ads_renderis listed. Nogooseworks:tools at all → say only: "Connect GooseWorks first — I make your ads through it." Tools but noads_render→ say: "Image ads can't be made on this connection yet." and stop. - Server start.
gooseworks:get_startedwithwants: "image"and the marketer's words; its answer settles setup and the brand only. A card → wait for the press (gooseworks:card_wait). - Research complete.
gooseworks:brand_read→research_statusmust becomplete, or every render fails after the yes.running→ wait (gooseworks:job_getkind: "brand_research");none→ say the brand isn't set up yet; no website or store link → images can't be made yet, so offer the concepts, templates and copy alone. - Load
references/engine-notes.md. - Pick the mode:
| Mode | Trigger | What runs |
|---|---|---|
| Fresh batch | new ads | Steps 1–12 |
| Iterate on a winner | "more like this ad" | Steps 1–12; the grid keeps the winner's reason and proof and changes one named lever per ad |
| Rotation | an open Rotation needed: for make-static-ads line in ads/campaigns/<slug>/rotation-requests.md, or pasted |
Steps 1–12, the reason taken from the request (the make-static-ads-concept-grid skill (../make-static-ads-concept-grid/SKILL.md) step 7) |
| Copy variants | "same image, different text / font" on an approved image | read the creative (gooseworks:ads_creative_read), then Step 11 |
| Rejected ad | "my ad got rejected", or an open Fix needed: for make-static-ads line in ads/campaigns/<slug>/rotation-requests.md |
the make-static-ads-fix-rejected skill (../make-static-ads-fix-rejected/SKILL.md) |
| Price or stock change | a price, offer or stock change on ads already delivered | list every delivered ad that shows the old fact (batch files; creatives by offer-ends and product tags) and flag each to pause; offer remakes with the new fact through Steps 6–10 on a new quote |
| Graphic request | a header, social or blog image, asked or from a sibling's Graphic needed: line |
the make-static-ads-graphic-request skill (../make-static-ads-graphic-request/SKILL.md) |
Steps
1. Read the brand, the brain, the account and what was made before
Load references/reading-the-brain.md. Reads: gooseworks:brand_read lean (summary, kit, products with product_ids, competitors, learnings; sections: ["organic"] for post topics from the last six months, giveaways skipped); the brand brain (gooseworks:brain_map, then gooseworks:brain_search for this product's claims and proof, §1); customer words, the angle bank, the performance read, rotation requests and campaign record (§2–5); gooseworks:meta_read view: "performance"; the brand's own Meta ads and its past GooseWorks ads (references/references-and-remix.md §2a).
Writes: a state note (winners, fatigue, the product's price and eligibility, the usable facts) and the concept inventory: each running own static in one line (reason, persona, offer, framework), plus the made-before list (reason, template, hook, example content shown).
Done when every source above is read or named as missing.
2. Ask the questions in one form
Reads: the state note, the kit. Writes: the brief and the answers. Use the host's question form (AskUserQuestion or its equivalent; in plain chat, one numbered message). Each question has a suggested answer first, so "use your suggestions" is a full reply. Ask:
- Logo: with the brand logo / no logo.
- Colours: brand colours / the template's colours / a fresh palette (different on every ad).
- Font: the brand font (offered only when it is on Google Fonts or has a metric clone,
references/fonts-and-color.md§1) / one of 2–3 fonts that suit the brand (§2) / no specific font. - Templates: you pick from a gallery of real ads running on Meta / I pick and show you what I chose.
- How many ads and where — suggested: 3 ads, Meta feed 4:5. Skip it when the ask already says. This is the playbook's only default count.
Also ask, only when it applies: a claim the concepts need that is not a usable fact ("can the ads say this?" — a yes is saved in the marketer's words through the save-learning skill (../save-learning/SKILL.md)); a sale price or offer on the product record ("show the $60 price?"); two sources that disagree on price or stock.
A physical product with no usable photos → stop: real product photos come first. Ask the marketer for them, or offer the goose-product-photos skill to make product shots first. Software → the kit's screenshots or real ones the marketer gives.
Done when every question has an answer.
3. Build the concept grid
Load references/creative-taxonomy.md. Reads: the brief, the concept inventory. Writes: the grid. Run the make-static-ads-concept-grid skill (../make-static-ads-concept-grid/SKILL.md). Done when every concept passed the concept gate and no two ads share reason + framework + style.
4. Show templates from the Meta Ad Library; the marketer picks
Reads: the grid, the Step 2 answers. Writes: the gallery and one template per ad. Run the make-static-ads-find-references skill (../make-static-ads-find-references/SKILL.md). Done when every ad has a registered template that the marketer picked (or the agent picked and showed), its concept unchanged, and structure-only yes/no recorded.
5. Write the copy around the image
Reads: the grid, voice, customer words. Writes: a copy row per ad. Run the make-static-ads-ad-copy skill (../make-static-ads-ad-copy/SKILL.md): primary texts, headline, description, CTA, destination, and the hook direction for the Variant brief. The on-image words are the composer's (Step 6). Done when every ad has a copy row.
6. Compose every ad as the server does
Reads: the grid, copy rows, templates, Step 2 answers. Writes: per ad and ratio, the final prompt and ordered image list. Run the make-static-ads-compose skill (../make-static-ads-compose/SKILL.md). Done when every ad × ratio passes the placeholder check (references/composer.md §7) and no image_urls holds another brand's template image.
7. Check claims and policy before any money is spent
Reads: copy rows, composed prompts. Writes: check results and the cut list. Run Pass A of the make-static-ads-claims-extract skill (../make-static-ads-claims-extract/SKILL.md) and the ad-policy-check skill (../ad-policy-check/SKILL.md). A cut on-image line → re-compose that ad (Step 6) and check it again. Done when every span is a usable fact or cut, and every ad has a claims line and a policy line.
8. Show the plan and the quotes
Reads: composed ads, check results. Writes: the plan table (references/handoff-format.md §3). One gooseworks:ads_render dry_run: true per ad and ratio (composer.md §10–11). Show each quote as returned and say: "If an image fails my check I redo it once, at the same price." Ask "Go ahead with these quotes?" Done when the marketer says yes.
9. Generate
Submit each call exactly as quoted. Poll gooseworks:job_get (kind: "ads_batch") until every pending is 0. Report a failed job; resubmit only on a new quote and yes. Done when every job is complete or reported.
10. Judge every render; retry the server's way
Reads: each render, its compose record. Writes: verdicts, retries, the drop list. Run the make-static-ads-render-review skill (../make-static-ads-render-review/SKILL.md): the blind judge on every render, one retry inside the yes, then the playbook's checks and the set-diversity check. Done when every render is passed, flagged needs_review, or dropped with its reason.
11. Copy variants — offered
For the one or two images the marketer likes most, offer the make-static-ads-copy-variants skill (../make-static-ads-copy-variants/SKILL.md): the same picture with other hooks or another font. They test copy, not concepts; each paid step gets its own yes.
12. Package and hand over
Load references/handoff-format.md; re-read the product record. Writes: lower-case tags on every creative (gooseworks:ads_creative_update, read-merge-write); the batch file ads/batches/<YYYY-MM-DD>-<slug>.md (gooseworks:file_write); a Rotation ready: answer under a served rotation request; the hand-off in chat in plain words, with each creative's link. Ask which to run. Offer a standing weekly refresh only when spend is steady and creative is fatiguing, created paused with gooseworks:automation_upsert (the spend-gate skill (../spend-gate/SKILL.md), automation rule).
Approval gates
- Spend — nothing renders before the marketer's yes to the quotes as GooseWorks gives them (the
spend-gateskill (../spend-gate/SKILL.md), branch A), never your own sum. That yes covers one retry per image at the same price, said up front. Paid ad-library reads: branch B, one yes to a ceiling. Layer splits and copy variants: each its own yes. - Questions and templates — the Step 2 form is answered before any quote; the marketer picks from the gallery or sees the agent's picks.
- Claims — every cut line is listed with what would confirm it. A claim the marketer confirms in this chat is saved in their words as a learning before it is used.
- The run decision — the human picks what runs; this playbook never launches or spends ad budget. Record a pick only on the human's words: feedback
rating: "happy"on the chosen render,pickedin its tags, and the pick in the batch file.
Output and write-back
- Tagged creatives in GooseWorks, registered templates, the batch file in the workspace, the hand-off in chat.
- Judge results go in the batch file and the creative's tags (
references/judge.md§8). Renderfeedbackcarries only the marketer's own ratings and words, because it trains the brand's learned rules. - Brand learnings only through the
save-learningskill (../save-learning/SKILL.md): the human's own words, one idea per learning, own superseded ones retired, text starting[ads YYYY-MM-DD](disapproval lessons, template lessons, look rules the marketer states — "no lo-fi", "always our serif"). Reasons to buy the founder states go in the batch file and the angle record. A claim is saved only as gate 3 says.
Done when
- Fresh batch: every concept passed the concept gate; no two ads share reason + framework + style, and with the text covered no two look alike; at least one concept the brand has not run before; every ad is built on a template the marketer saw or the agent picked and showed; the picks are recorded.
- Every ad: composed with the marketer's words as the only User instructions; judged blind; at most one retry; passed, flagged or dropped with its reason; every line, including text drawn in the image, a usable fact; tagged and in the batch file.
- Iterate / rotation: one named lever changed per ad from the winner; rotation answered under its request.
- Copy variants: same picture, only the words or font changed, each line through the claims check.
- Rejected ad: the violating line or element is gone, both checks pass, and
Fix ready:is answered. - Price or stock change: every affected ad is listed with a pause flag; remakes only on a yes.
Out of scope
- Launching, audiences, budgets, kill/scale, scanning the account for rejections — run-paid-campaigns (this playbook recommends a test structure and fixes rejected image ads).
- Search text ads (Google / Bing RSA headlines, keywords) — not covered yet.
- Which ad won — analyze-performance. Video ads — goose-video. Product photography — goose-product-photos (GooseWorks skills, not playbooks).
- A full competitor teardown and the angle bank — research-competitors (this playbook reads live ads only to find templates and reasons to buy); customer language — research-market.
- Landing pages — analyze-performance's page audit (this playbook flags a destination that does not match).
- Approving claims in bulk — not covered yet; the lines this batch needs go through the claims-gate skill. Retail-media creative for marketplaces; decks, print, packaging, multi-page layouts.