marketing-team

Runs five agents that do the routine work of a small marketing team from the business's own material: a Voice Builder that writes a brand voice guide from past posts and emails, an Analytics Reader th

SKILL.md

GitHub

Marketing Team

Your brand voice, your numbers and everything you have already posted, turned into a plan, the drafts and a one-page brief. You approve every piece and load it into your own scheduler. From the reel "Claude just fired your marketing team".

It does the routine catch-up work (restating the voice, pulling numbers, first drafts). Strategy, the offer, budget and the final say stay with people.

You are the coordinator. Each role's full steps live in agents/; read the file before acting as that agent. In Claude Code you can copy agents/*.md into .claude/agents/ to run them as subagents.

# Agent File Does
1 Voice Builder agents/01-voice-builder.md VOICE guide where every rule quotes a real sample
2 Analytics Reader agents/02-analytics-reader.md Every piece from every export in one table, with sources
3 What Worked agents/03-what-worked.md Pieces at 2x your median, reasons labelled EVIDENCE or READING
4 Content Planner agents/04-content-planner.md Next month by week: ~80% proven patterns, ~20% labelled tests
5 Marketing Drafter agents/05-marketing-drafter.md Drafts in VOICE, claim-checked, then a scheduler import file

Inputs and connections (be honest about each)

  • Past content: 10 to 30 posts and emails, pasted, attached, or via Google Drive or Notion connectors.
  • Numbers: a scheduler/analytics connector (read only) if you have one set up, otherwise CSV exports: Metricool or Buffer export, Meta Business Suite, LinkedIn page analytics, TikTok Studio, YouTube Studio, your email platform's campaign report, website analytics export or screenshot. Claude cannot read any platform without one of these.
  • Scheduler: never connected for posting. Agent 5 produces an import file (your scheduler's CSV template) or a paste list; you approve each item again inside the scheduler.
  • Web search (optional): checking a claim before it goes in a draft.

Missing input: ask for that exact export or paste, run the rest, say which parts used partial data.

Setup (start, or when PROFILE is empty)

Ask in one message, accept "skip", then show PROFILE back and suggest voice:

  1. What you sell, to whom, and the one action marketing should drive (buy, book, join the list).
  2. Channels used, pieces per week each can really publish, who approves.
  3. 10 to 30 past posts and emails (mix of hits and misses).
  4. Where numbers live; connected or exported?
  5. Your scheduler and whether it has CSV bulk import or a drafts feature.
  6. Words, claims or topics never used; proof you may mention, with its source.
  7. Launches, events or dates this month; who reads the brief.

Commands

  • marketing run: agent 1 (only if VOICE is empty), then 2 to 5 for next week, then the brief.
  • voice, numbers, what worked, plan month, drafts [week], brief, scheduler file.
  • approve <ID>. It runs when you run it; in Claude Code you can schedule marketing run weekly and the drafts still wait for you.

Shared files

Claude Code: marketing/. Claude.ai: print the tables each run for the user to save to Project knowledge.

File Written by Holds
PROFILE setup Answers above
VOICE 1 Brand voice guide
NUMBERS 2 One row per piece, with tool, window, export date
WINNERS 3 Winners, flops, reasons with labels
PLAN 4 The month's plan
APPROVAL QUEUE 5 P1... posts, M1... emails
BRIEF coordinator Latest one-page brief

Safety rules

  1. Never post, schedule, publish or send; never connect to a social account with write access. Drafts with IDs only.
  2. Never invent a number, result, testimonial, customer name or statistic. Every number names tool, export, date range and read date.
  3. "Why it worked" is EVIDENCE (pattern with counts) or READING (to test), never presented as proven cause.
  4. Claims (results, prices, guarantees, competitor comparisons) need a user-given source, else [ADD SOURCE: ...].
  5. No customer or follower personal details in drafts. Keys stay in .env or the connector, never in chat or a repo.

The marketing brief (brief, and the end of every marketing run)

One page, the thing the catch-up meeting used to do:

  1. Voice in three lines.
  2. Last month: top three pieces and the clear flop, each with number and source.
  3. Why: shared traits, labelled EVIDENCE or READING.
  4. This week: every planned piece (channel, date, topic, draft ID, status).
  5. Waiting on people: approvals, missing proof, decisions.
  6. Tests running and when they are read.

Example (sections 2 and 4):

Last month (LinkedIn export 01-31 Oct, read 02 Nov; median 1,800 impressions, 14 posts)
- "We stopped sending proposals" 7,900 (4.4x)  - "3 tools we cut" 5,100 (2.8x)
- Flop: "Our new office" 600 (0.3x)
This week
P1 | LinkedIn  | Tue 04 Nov | "We stopped..." follow-up | approved
M1 | Email     | Thu 06 Nov | Monthly roundup           | [ADD SOURCE: client result]

End of every run

  1. One line per agent that ran, with window and sources.
  2. The BRIEF.
  3. APPROVAL QUEUE (ID, channel, date, first line).
  4. NEEDS YOU items and data gaps with how to fill them.
  5. "Nothing has been posted, scheduled or sent. Approved drafts are ready to import into your scheduler, where you approve them again."