How to Build an AI Marketing Team in Claude That Runs Without You Every Week

You already built one of these. A custom GPT, a project, a Claude agent. It demoed well once. Week 1 you used it every day. Week 2 you noticed you were editing everything it gave you, and editing before you even read it. Week 3 you skipped it, because doing it yourself was faster. Nobody could tell you why it degraded, so you concluded the tool was not there yet.

The cause is almost always the same. You gave one agent 5 jobs. Research, writing, reporting, campaign management, and answering your questions on top. "A thing with 5 jobs does all 5 about 4 out of 10," Will Leatherman said in the workshop, "which is worse than useless, because now you have to check all 5 every time."

An AI marketing team is a set of small agents, one job each, each defined by a plain markdown file in its own folder, that read your data, write against your examples, hand work to the next agent, and run on a schedule. It is built for the marketing lead on a small team who is effectively the whole marketing department. The version below runs a real bootstrapped agency, has been in production since April, and was rebuilt live from an empty folder in a single session.

What is the difference between a chatbot and an agent?

One answers, the other finishes. "ChatGPT answers you, and then you do the work," Will said. "An agent does the work for you. It opens your files, it runs the searches, it creates the documents, and it keeps going until the job's done."

Chat tools do tasks. Your job is made of jobs. The numbers, the research, the content, the campaigns, every single week, whether or not you have the time that week. A team is what covers the weeks you do not.

What are the 4 rules that keep an agent alive?

Every decision in the build traces back to one of 4 rules. Skipping any one of them costs a full rebuild later, which is what happened to the agent you already abandoned.

What goes in the CLAUDE.md file every agent inherits?

Start with an empty local folder and one file named CLAUDE.md, in capitals, with the `.md` extension. The name matters because that is the file the session reads first when it enters the folder, which Claude Code's memory docs describe as the project context every run picks up. The build in the workshop runs in Claude Cowork against a plain folder on a laptop.

One warning before you pick the folder. Do not build this inside an iCloud folder on a Mac. Will has seen repeated failures where files were not synced locally and writes broke silently.

The file itself is plain English in 4 short sections.

1. Who you sell to. In this build, B2B companies between 50 and 500 people, buyer is a VP of marketing.
2. What you sound like. Direct, no hype, numbers instead of adjectives, shorter sentences.
3. The channels you run.
4. The numbers that matter to you.

Write it once and every agent you add inherits it. Skip it and the agents you add in month 2 lose the context, start hallucinating your business back at you, and whoever you onboard to run them will not catch it.

How do you teach an agent what good looks like?

Point at a real example rather than describing one. After the job description, the Analyst knows what to do and has no taste. It does not know what separates a good report from a bad one, and left there it will write something that reads like AI slop with your own numbers in it.

The fix takes about 90 seconds. Drop in a report you were actually happy with and ask the agent to read it, annotate it, and work out what made it good.

"I'm not writing a style guide. I'm not describing what good is. I'm just pointing at what good actually looks like."

This is also the answer to the prompt engineering question that comes up in every one of these sessions. Reliability comes from codified files the agent reads on every run. The agent file itself is 3 lines. The job, which is read the numbers in the marketing sheet every Monday morning. What it produces, which is a short report covering what moved, why it moved, and one action per number. Where it goes, which is the marketing Slack channel.

Which data source should the first agent read?

One spreadsheet. A CSV with a column for the week, a column for the channel, and the metrics that matter, so demos booked, pipeline generated, reply rate.

That looks boring next to a CRM connector, and it is deliberate. "Every source you add is another thing that can quietly break," Will said. You need roughly a month of the report being right before you stop second guessing it, and each extra integration pushes that date out. Teams outside marketing ops are also frequently not allowed to wire up MCP servers or API keys, so a shared sheet is the honest path for most people.

One trap surprised the team enough that it became a standing line in every agent file. The freshest data is usually the wrongest data, because ad platforms and analytics tools revise yesterday's numbers overnight. An agent reporting on yesterday will hand you a confident recommendation that normalizes by the next morning.

The fix is a single line in the file telling it to read a window that ends 2 days ago. You lose 2 days of freshness and you stop getting false alarms.

Wiring it up is one instruction, and it is worth copying the shape of it.

"Add one line to the analyst markdown file pointing to this CSV, and one line telling it to read the example report before writing rather than describing the shape. Do not change anything else in the file."

If you are further along and connecting a real data tool, the instruction is identical. Swap the filename for your connector.

What did the live run actually produce?

A 4 minute run and a report with a flagged gap in it, which is the point. Pipeline came in at 240,000, up 9% from 220,000. LinkedIn added 35,000. Email lost 15,000. Demos booked were flat. The total rose while one of the 2 channels fell, and every email number dropped in the same week every LinkedIn number rose, which the report called out as either one cause or two rather than picking one.

It also flagged that the data set it had been given was too thin to be sure. That is rule 4 working in public.

The second run surfaced a real defect. The report had no date range at the top, so there was no way to check which week it described. The correction was one sentence of feedback into the file. Build the thing stage by stage and you find that defect in 30 seconds. One shot the whole workflow and you are debugging 5 steps at once with no idea which one failed.

How do you put the agent on a schedule?

You say "schedule this Monday at 7am every week." That is the entire step, and almost nobody does it, which is the actual reason most people do not have agents.

Monday at 7am is a deliberate choice. It lands before the first meeting of the week, so there is time to read the report before anyone asks you about it. Check your source refresh before you lock the time. If the spreadsheet only updates on Tuesday, a Monday run reports last week twice.

"This agent is something businesses pay real monthly retainers for, and we built that whole thing in about 20 minutes."

Scheduling is also what unlocks agents working together, because a report that arrives reliably is a report the next agent can read. Slack's incoming webhooks are the simplest destination for the output.

What do the Researcher, Writer, and Distributor do?

Each one takes the previous agent's output as its input, which is the same one job rule applied across a chain rather than inside a file. Claude Code's subagents formalize the same split.

The Researcher is the one most likely to fool you. Ask any agent what AI says about your category and it will answer fluently from its own memory, which is stale data from a different system than the one your buyer is using, and it cannot tell you that. So the Researcher quotes what an engine actually repeats back, records the question it asked, surfaces the date, and asks 3 times to report the spread. If you want the same discipline as a standalone habit, the 20 minute AI search ranking audit covers the manual version.

The Distributor exists because one idea should travel. "Effective marketing is saying the same thing as many different ways as you can." Cutting a single draft into a week of platform native assets is a whole build on its own, covered in the Claude workflow that turns 1 daily idea into 5 ready to post drafts.

What should always need a human to say yes?

Two lists, written into the same context file, so every agent you ever add inherits them.

The agents handle these unsupervised. Read, draft, report, pull numbers, watch competitors.

A human signs off on these every time. Publishing, sending, spending money, and anything that lands in front of another person.

In Will's agency this is a written policy enforced in the GitHub repo, not a preference. "Nothing goes out under a client's name without someone saying yes."

This is also the part that gets the system approved internally. A governance layer your boss can read in 10 seconds is the difference between a shadow tool and a sanctioned one. The AI marketing ops build for CMOs goes deeper on the operating structure, and how GTM teams scale with AI agents covers the same approval question across sales and marketing together.

The takeaway

The finished team is 7 markdown files. A context file, 4 agent files, 1 example report, 1 data connection. No software to maintain, and no new line item, since it runs on the Claude, Slack, and Notion subscriptions you already pay for.

One honest limit, stated in the workshop. There is no measured before and after on hours saved, because every team's baseline is different. What is measurable is that the Analyst ran this morning, took 4 minutes, and posted a report nobody asked it for.

Give yourself a week. Write the context file and the Analyst, run it for 45 minutes until the numbers are right, then add the Researcher and wire the first handoff. By next Monday you will have a full brief landing before your first meeting.

Once your team is shipping, the next question is whether AI search can see any of it. Run the free AEO audit to find out where you currently rank and what to fix first.