Build a Small AI Marketing Team with Claude Code Subagents

Every few weeks someone shows me a diagram of eight AI agents passing work to each other like a factory line. It looks impressive. In practice, most of those setups break at the handoffs, because nobody decided what each agent should hand to the next one. This guide shows you how to build a small marketing team in Claude Code that actually holds together, starting with four roles and a folder of plain files.

Start with the output, then design the roles

Before you create any agent, write down the one thing this team should produce every week. A launch email and three social posts. A landing page draft plus ad copy to test. A monthly content plan. Pick one.

Then work backwards. Who needs to find the facts? Who decides the angle? Who writes? Who checks? For most small teams that is four roles: a researcher, a strategist, a writer and a reviewer. You can add more later. Starting with eight is how you end up with eight mediocre agents instead of four good ones.

Set up the project folder

Make a folder for your marketing work and open Claude Code inside it. The structure below is only a suggestion; Claude Code does not require it. What matters is that every role knows where to read and where to write.

marketing/
  CLAUDE.md
  brand/
    audience.md
    voice-samples.md
    offer.md
  research/
  strategy/
  drafts/
  reviews/

CLAUDE.md is the file Claude Code loads at the start of every session in that folder. Keep it short: who you sell to, where the brand files live, and your rules. Put the long material in the brand/ files and tell Claude when to open them. The memory docs explain how this file is loaded.

# Marketing workspace

We sell [product] to [audience]. Brand files live in brand/.

Rules
- Every factual claim needs a source in research/.
- Never invent customer quotes, numbers or results.
- Nothing gets published, scheduled or sent without my approval.
- Write in the voice shown in brand/voice-samples.md.

The voice samples file does more work than any style description. Paste three to five pieces you wrote and liked. Real text beats adjectives like "friendly but professional" every time.

Create the four roles as subagents

In Claude Code, a subagent is a Markdown file with a short header, saved in .claude/agents/ inside your project. Each one runs in its own context, so the reviewer does not inherit the writer's reasoning. You can write these files yourself or ask Claude to create them. The subagents docs list every field.

Here is the researcher. Notice the tools line: it can read files, search the web and save its notes, and nothing else.

---
name: researcher
description: Finds and verifies facts for marketing work. Use before any strategy or draft.
tools: Read, Grep, Glob, WebSearch, WebFetch, Write
---

You research for a small marketing team.
Read brand/audience.md first.
For each claim we might make, record: the claim, the source URL,
the exact supporting detail, and any caveat.
Mark anything you could not verify as UNVERIFIED.
Save your work to research/[topic].md. Do not write copy.

Make the other three the same way:

• strategist reads the research and the offer, then writes one page in strategy/: the angle, the main message, three supporting points and what to avoid saying.
• writer reads the strategy and the voice samples, then writes drafts in drafts/. It may only use claims that exist in the research file.
• reviewer reads the draft next to the research and the brand files, then lists exact lines to change and why, in reviews/. It does not rewrite the whole thing.

Give each role only the tools it needs. The writer does not need web access. The reviewer does not need to write drafts. Tight permissions keep each role honest about its job.

Run the handoff from the main session

You do not need a separate orchestration tool. Your main Claude Code session is the coordinator. Paste something like this:

We are producing a launch email and three LinkedIn posts for [offer].
1. Use the researcher agent on [topic]. Wait for research/[topic].md.
2. Use the strategist agent on that file.
3. Use the writer agent to draft from the strategy.
4. Use the reviewer agent on the drafts.
5. Apply the reviewer's line edits once, then stop.
Show me the final drafts and list every UNVERIFIED item.
Do not publish or schedule anything.

The single revision cycle matters. Without a limit, a writer and a reviewer will happily rewrite each other forever, and the text gets blander each round.

Keep yourself in the loop where it counts

Claude is very good at research synthesis, positioning, copy and reading numbers you give it. It should not be the one pressing publish or spending money. I keep three gates for myself: approving the strategy, approving the final copy, and anything that leaves the building.

Every time you rewrite a line the writer produced, add the before and after to voice-samples.md with one sentence on why. After a few weeks the drafts need far fewer edits, because the team is learning from your actual corrections instead of your description of yourself.

Want to see a finished example first?

If you learn faster by reading a working setup, look at ai-marketing-claude, an MIT licensed pack of marketing skills and five audit agents for Claude Code. Its /market audit command runs the agents in parallel on a website and writes a report. Clone the repo and read install.sh before running it, and read the scores as an opinion scored against a rubric. Two runs on the same site can land on different numbers.

I find the most value in reading how the agent files are written, then building my own smaller version around the output I actually need.

No terminal? Use Projects

If you work in the Claude app, Projects give you the same idea with less setup. Create one project per role, paste the role instructions into the project instructions, and upload your brand files. You move the output from one project to the next by hand. It is slower, but you see every handoff, which is a good way to learn what each role really needs.

Start this week with two roles, a researcher and a writer, and one real piece of content. When the handoff between them works without you fixing it, add the reviewer. Then the strategist. The team grows only when the previous step is boring and reliable.