How to Build a 4 Agent Marketing Team in Claude Cowork That Turns Last Week's Numbers Into This Week's Content

Workday surveyed 3,200 full time employees who actively use AI and found that 37% of the time AI saves goes straight back into rework, so correcting, clarifying and rewriting what the AI produced. Only 14% said they get a clear win out of it consistently. The work did not disappear. It moved from doing to checking, and checking is the slower half when you did not do the work yourself.

"The work didn't actually go away. It moved from doing to checking. And checking is slow when you didn't do the work yourself." — Will Leatherman, Founder of Catalyst

That is the whole design constraint. Every agent below is shaped so you can verify what comes back faster than you could have done the job. Catalyst has run its own content on 7 agents since April, across 150 B2B companies and more than $14M in pipeline from content, and this 4 agent version is the smallest useful shape of it.

What is a 4 agent marketing team in Claude Cowork?

It is 4 plain text files in 1 folder, each describing 1 job, that hand work to each other in a line and run on a schedule without you starting them. Claude Cowork is Claude working out of a folder on your computer, so the work is still there next time you open it and it can run on a timer. An agent is a text file with a list of steps, written the way you would write an SOP or a job description. A handoff is 1 agent leaving a file where the next one looks for it.

Read the arrows, because they are the point. Your numbers decide what the team makes that week. Most content agents open on a blank page and a topic somebody picked back in March. This one starts from what actually worked last week.

"If you can write a job description for a new hire, you can write every file you're about to see." — Will Leatherman, Founder of Catalyst

Why does 1 job per agent matter so much?

Because a failure in a 4 job agent has 4 places to hide. The first thing most people build is 1 big prompt or a custom GPT with everything stuffed in it. The output comes back long, the mistake could be anywhere in it, and now you are reading every word every time.

"The second you give one agent research and writing and reporting, it does all three badly. And now you have to go check all three and you can't find the failure point." — Will Leatherman, Founder of Catalyst

You would not hire 1 person to be your analyst, your researcher, your writer and your social lead. 4 short files are 4 quick reads. 1 long file takes the afternoon. Catalyst's own 7 agents split the same way, so 1 collects signals from social, 1 routes each signal to the right client, 1 scores it, 1 picks the best of the week, 1 drafts the post, 1 checks the draft against the rules and 1 publishes. When one of them broke, we knew which one it was.

What goes on an agent's 5 line card?

Every job on the map fits on 1 page with 5 lines.

1. Who it is for.
2. The job in 1 line.
3. What it takes in.
4. What it hands off, and to whom.
5. The 1 thing it should never do.

The last line is the one people skip. Every agent gets exactly 1 never line, and that is what makes a folder of text files read like a team you can rely on. The analyst's card is for the marketing lead who gets asked about the numbers every week, its job is to read the numbers and write the weekly pulse, it takes 1 exported sheet, it hands the pulse to the researcher, and it never names a cause the sheet cannot show.

What belongs in the CLAUDE.md file every agent reads first?

Four things, written in sentences. Who you sell to, what you sound like, which channels you run, and which numbers matter. Cowork reads this file before any agent does anything, which is why it is the file that stops every draft sounding like it was written for somebody else's company.

Here it is for a B2B company of 50 to 500 people. The buyer is a VP of marketing with no team under them. Direct, no hype, numbers instead of adjectives. LinkedIn, email and the website. Pipeline, demos booked and reply rate.

Write it in sentences rather than bullets, because a bullet that says "VP of marketing" drops the reason. A sentence saying they have no team under them tells the agent why that person behaves the way they do, and that is what changes how it writes. Run 1 test on every line you add. Could an agent behave differently because of it? "50 to 500 people" passes, because it changes who the agent writes for. "We value quality" fails, so it stays out.

The line doing the most work is what you sound like. "Numbers instead of adjectives" is 4 words, and it is the instruction that stops every output reading like a press release, because every agent in the folder inherits it.

The bottom of the file is 2 lists, and they answer the question your boss asks before any of this goes near a customer. What is it allowed to do without asking you? Reading, drafting, reporting, pulling numbers and watching competitors run alone. Publishing, sending and spending always wait for a person. Drafting can be undone. Publishing cannot.

A CMO posted what happens when nobody writes those 2 lists down. His team handed an agent a small job, so add 10 new banner images to 1 ad, take the old ones out, leave the videos alone. Four minutes later it had deleted all 19 videos, deleted half the old images, and added none of the new ones, on the ad carrying most of their spend. The deletion had an undo button. The expensive part was the next day, when the team believed the work was done and planned around it until somebody read the change log line by line.

"Never hand an agent anything you can't verify faster than you could have done yourself." — Will Leatherman, Founder of Catalyst

The 2 lists are a map of where your checking happens.

Why build the analyst first?

Because your boss reads the numbers, and nobody on your board is asking whether you posted 4 times last week. They are asking what happened to the number you just reported. The analyst is also the agent that tells the other 3 what worked, so the boring one makes the fun ones better.

Its whole job file is 4 lines. The job, what to produce, where the output goes, and when it runs. No role play, no telling it that it is a world class analyst. If a competent person could do the job from what is on that page, so can the agent. If a person would come back with questions, those questions are exactly what the file is missing.

What makes the pulse sound like a person is 1 real example with notes next to it, and that example has 5 parts.

An action like "improve your email strategy" is useless, because on a busy morning nobody knows what to do first with it. And the "I do not know" line is the one that earns its place, because AI skews confident. Without it you get a cause the model cannot possibly know, dressed up as an insight, which is exactly why it gets past you.

When you build your own, paste in the last weekly update you actually sent, flaws and all. Then when a pulse comes out wrong later, do not fix the pulse.

"Fixing the pulse fixes one week. And fixing the example fixes every week after that." — Will Leatherman, Founder of Catalyst

Keep it to 1 data source for the first month. Every source you add is another thing that can quietly break and another thing you have to check. Earn the fancy connections after 4 accurate pulses in a row.

How would you know if an agent stopped working 3 weeks ago?

You would not, unless you built the alarm in. A silently failing agent produces a pulse that looks exactly like a good one. The numbers are real, the sentences are clean, and nothing on screen says a week went missing.

"One of our agents logged success for six weeks in a row while producing nothing. Another one ran for seven weeks reading none of the material it was supposed to judge against." — Will Leatherman, Founder of Catalyst

The fix most people reach for is a second agent to check the first one. A head of marketing built exactly that in September, an agent whose only job was to catch her dashboard lying, and then it got one wrong and sent her hunting for leads that were never missing. Checking got cheaper and it still had to happen.

The real fix is 2 lines inside the job itself.

1. Print the date range you actually read at the top of every pulse. If it opens with 2 weeks that are a week old, you catch it in 1 second.
2. Write the pulse every week even with nothing to say, so "I ran and nothing moved." Once it always shows up, a missing pulse is the alarm.

Both of the failures above would have been caught by those 2 lines, and each takes 1 second to read.

What did the team actually produce on real numbers?

The workshop ran the folder live on a sample sheet. Here is what came back.

The flat number is the interesting one. Demos booked did not move, because 2 channels moved in opposite directions and cancelled out. A dashboard shows you a flat line and you move on. The pulse showed the 2 stories underneath it, then said it did not know why LinkedIn climbed, because nothing in its sheet splits LinkedIn by post.

The researcher picked up that exact gap. LinkedIn went from 12 demos to 17, and it found that 1 post, "We cut our own weekly reporting from 3 hours to 10 minutes," booked 11 of those 17. It named the post, the demos and the sheet it read them from, so checking the claim means finding 1 row. Why email fell from 8 to 3 it left for a person.

The writer's draft came back with 2 blanks in square brackets where a before number and an after number belong, because the rundown did not have them.

"That gap is the most important thing on the screen." — Will Leatherman, Founder of Catalyst

You fill in 2 blanks and trust the rest, which is about 10 minutes. The alternative is an agent that guesses and a human who reads every sentence hunting for the invented one. The distributor was asked for 5 social posts and wrote 3, and said why, because the draft only had 3 lines that stand on their own. It also produced 3 emails and a one pager for sales, with every line traceable back to the draft. That same handoff logic is what powers a full transcript to a week of platform native content once you trust the drafts.

The writer does no research on purpose.

"An agent that can go and find things will, and then it writes about whatever it found." — Will Leatherman, Founder of Catalyst

And the distributor writes no new claim, because a repurposed post that says something new is a post nobody reviewed.

What do you fix when an output comes back wrong?

The file that caused it, never the output. Fixing the output fixes 1 version. Fixing the file fixes every week after it. Three show up first.

If something goes badly wrong, pause the scheduled task rather than deleting it. Fix the 1 file, run it once, read it, turn it back on. That is a 5 minute repair when the agents are this small.

How much you let an agent do moves in 3 steps. Step 1 it reports what it found and stops. Step 2 it drafts the change and waits for you. Step 3 it makes the change itself. Everything in this build sits on steps 1 and 2 on purpose, and plenty of teams will never move anything customer facing to step 3, which is fine. One habit helps between agents. Ask each one to score its own confidence from 1 to 10 and tell the next agent to take only what scores 8 or higher, so a 4 scored on 1 week of data stops there instead of becoming a post.

Two places these teams get stuck. The input, where you are still dropping a sheet in by hand, which is fine for the first month and worth connecting directly once the pulse has been right for 4 weeks. And the output, which is the one that kills most of them, where the files sit in a folder and nobody does anything with them. Every output ends on 1 action for 1 person, delivered where that person already looks for their tasks. The last thing anyone needs is another report. Teams running this at scale, like Jacob Bank's 9 person team with 300 agents, solve the output problem before they add agent number 5.

How do you grade your own AI setup?

Score your current setup green, yellow or red on all 5. Yellow means partly true. Each red points at exactly 1 file.

If 1 agent does the research and the writing, that is a yellow on 1. If you are not sure what it is allowed to send, yellow on 2. If you have ever caught it inventing a reason, red on 3. If only it runs when you remember, red on 5, which is the most common red there is.

Mostly red? Start with 1 and 3, because those are the 2 that make checking fast. The others only matter once there is something worth running every week. None of this adds a subscription either, since it runs on Claude in a folder of plain text files, so month 2 costs the same as month 1.

What should you build in the first month?

One agent a week.

Week 1 is 4 steps you can finish in one evening. Open the free 4 Agent Marketing Team Starter Kit as a folder in Cowork, paste your last real weekly update in as the example, drop your own numbers into the spreadsheet or keep the sample ones, then run the analyst once and read the pulse out loud line by line. Run it start to finish manually before you automate it, because a first run can stop halfway to ask permission and a scheduled run that stops just sits there saying nothing. Schedule it for 7am so the pulse lands before the first meeting where somebody asks about the numbers. A pulse that shows up after that conversation is just another file. And leave the computer awake, because the folder lives on your machine.

"You can't remember to use a tool every week and you won't remember to run this every week." — Will Leatherman, Founder of Catalyst

Add the researcher when 3 things are true at once. The pulse has been accurate several weeks running, reading it takes you a few minutes, and the date range line has never surprised you. Then run the same test on the researcher before the writer goes in.

Messy data is the most common reason people stall here, and it is a bad reason. The analyst runs off a single spreadsheet on purpose, and the date range line means it shows you which week it read rather than writing around the gap. You also do not need Claude Code for any of it, since everything here is Cowork and Claude's own scheduled tasks working on a folder of text files. The same files run in Claude Code if you already use it, and nothing in the kit needs it.

The takeaway

Build the analyst this week and put it on a 7am schedule, because the other 3 agents are variations on a file you already understand. Before you hand any agent a job, run the test.

"Before you hand an agent an actual job, ask whether you can check what comes back faster than you could do it yourself." — Will Leatherman, Founder of Catalyst

The 5 checks are how that answer becomes yes. One job each, you know what waits for you, it says when it does not know, it tells you when it broke, and it runs on a schedule.

If you want the step by step version of the same build from an empty folder, how to build an AI marketing team in Claude walks the first agent end to end, and the AI agents playbook covers which repetitive marketing tasks to hand over first. Once this team is publishing every week, run the free AEO audit to see what ChatGPT, Claude, Gemini, Perplexity and Grok actually tell buyers about you. It takes a work email and about 5 minutes, and what it finds is what the team should be writing about next.