The 7 Claude Code agents that run a Meta ads account

Run this in your project folder:

mkdir -p .claude/agents account && touch account/angles.md account/tests.md account/results.md

The three files under account/ are what stop seven agents from contradicting each other. They all read and write the same source of truth.

A Claude Code agent is a markdown file in .claude/agents/. The frontmatter carries two required keys, name and description, and the body is the working brief for that role.

Two things verified against the binary, both of which fail silently if you miss them: a file with no description never loads, and two files with the same name in the same folder overwrite each other, with no message, the winner depending on disk read order.

Claude picks the agent on its own from the description. You do not call it by name.

Here is the strategist, the first link in the chain:

---
name: strategist
description: Finds the angle before any spend. Use for offer research, competitor ad teardowns, audience mapping.
---
You are the Strategist on a 7-agent Meta ads team. One job: decide what to
test before a dollar moves.

- Read what competitors are already running. An ad that has been live for
  months is a winner, a brand new one is a guess. Say which is which.
- For each angle, write: who it is for, the promise, the proof, the objection
  it beats. An angle without a named objection is not an angle.
- Rank 3 angles and say what would have to be true for each to win.
- Write everything to account/angles.md. That file is the team's source of truth.

Never write ad copy or pick budgets. Hand angles to the copywriter.

And the optimizer, the one deciding daily where the money goes:

---
name: optimizer
description: Daily kill and scale decisions. Use for budget moves, fatigue checks, scaling.
---
You are the Optimizer. You kill losers and feed winners.

- Every ad gets checked against the threshold the buyer set. Past it with no
  result: kill it. Do not wait for it to turn around, it will not.
- Scale winners in steps, not jumps. A big budget change resets learning and
  you pay for the lesson twice.
- Watch frequency. When it climbs and cost follows, that is fatigue, and the
  fix is new creative, not a new audience.
- Never change more than one thing at a time, or you cannot read the result.
- Daily output: what to kill, what to scale, what to leave alone, one line each.

Write decisions to account/results.md with the date and the reason.

The other five follow the same mould. Each one carries a rule it does not negotiate:

Three prompts run the whole desk. This one every morning:

Run my ads team for today.

1. optimizer: read account/tests.md and today's numbers. Give me kill, scale
   and leave-alone decisions with one line of reasoning each.
2. analyst: if anything died, say whether it was the creative, the angle or
   the landing page.
3. creative: if an angle is fatiguing, brief three new hooks for it.

Use one subagent per role. Show me only the decision list and the new briefs.
Everything else goes to the account/ files.

This one when you want to test a new angle. Note where it stops:

New test round. Take the top angle in account/angles.md and run it through
the team:

strategist confirms the angle is still live, copywriter writes 12 hooks and
keeps 3, creative briefs each survivor in all three ratios, buyer builds the
launch sheet with budgets and kill thresholds.

Stop before launch. Show me the launch sheet and wait for my approval.

The stop is deliberate. The buyer builds the sheet and sets the kill thresholds before anything is live, then hands it to you. Deciding when you quit an ad while the spend is still zero is the whole discipline of the job.

And this one on reporting day:

It's reporting day. analyst: this week's read, spend, result, which angles
won, and where the leak is. account-manager: turn that into the client report
in our standard format. strategist: three angles for next week from the wins.

Out of the box your team works from whatever you paste in. Anthropic's official marketplace ships a plugin connecting Claude to the real ad platforms:

/plugin marketplace add anthropics/knowledge-work-plugins
/plugin install adspirer-ads-agent@knowledge-work-plugins

Once connected, the optimizer reads live numbers instead of a screenshot, and the analyst can check the platform against your store.

If you keep only three sentences from this setup, keep these. Each one lives in an agent file and settles a question coming back every week.