An AI CMO without a plan is just a very fast intern
What an AI marketing stack really takes off a CMO, and what it cannot. The open source Claude Code plugin I built, and the two files it needs first.
GTM Architect & Growth Operator · Now · 28 August 2026
TL;DR · Key insights
- AI can take the production off a CMO. It cannot do the deciding, because nobody ever wrote the decisions down
- Write two files: who you sell to, and what you are betting on this quarter. Everything else reads them first
- A request that matches neither bet gets a written refusal with a price on it, not a plan
- Cap production at what you can genuinely review, or you become the bottleneck you automated around
- Operator CMO is my open source plugin for Claude Code. Nine skills, MIT, two lines to install
Ask an AI to write your landing page and you will have one in ninety seconds. Well structured, correctly spelled, confident.
It will also be about a company nobody described, aimed at a buyer nobody picked, making a claim nobody checked.
That is not a writing problem. Nobody ever told it who you sell to.
The plugin
What an AI marketing stack can and cannot replace
Most AI marketing stacks are a folder of skills: one writes blog posts, one does social, one does SEO, one does email. They work. What they replace is a marketing operations team, and a fast one.
They do not replace a CMO, because almost nothing a CMO gets paid for is production.
| The work | A CMO | Operator CMO |
|---|---|---|
| Choosing which segment gets the quarter | Decides | Refuses to guess. Asks, and records your answer |
| Killing a campaign people are attached to | Decides, and takes the heat | Fires the stop date you set and shows you the change |
| Saying no to the founder's podcast idea | Decides, and explains the cost | Writes down why it fits nothing, with the price of changing the plan |
| Judging whether a claim is true | Decides | Will not use a claim you have not evidenced |
| Writing the page, the email, the brief | Hands it to someone | Does it |
| Building and filtering the target list | Hands it to someone | Does it, applying your rules literally |
| Getting found in search and quoted by AI | Hands it to someone | Does it |
| The weekly numbers | Hands it to someone | Does it, and flags what changed |
| Noticing the plan stopped being true | Decides | Spots the drift, proposes the edit, waits for a yes |
The left column is the job. The middle is what a CMO does with it. The right is what the plugin does. Notice which side the deciding stays on.
The top four rows are the job. The middle four are what you were paying somebody else to do anyway. The gap in most stacks is that nothing holds the top four, so every skill quietly invents its own version of the strategy each time it runs.
A stack with nothing written down does not produce bad writing. It produces good writing about nothing anyone decided to say. That looks like productivity for a quarter, which is exactly how long it takes to notice none of it added up.
What you have to write down first
Meet Ada. She sells software that shortens month-end close for manufacturers with an ageing ERP, 200 to 2,000 employees, in Germany, Austria, Switzerland and Poland. Her marketing team is her, two contractors, and a folder of AI skills.
Before any of those skills writes a word, it reads two files kept in the same repository as the work.
Who she sells to. Not just company size, but the trigger: an ERP upgrade, a plant acquisition, or a failed audit in the last six months. Without one of those, the problem is real and the budget is not. Plus who she turns away, in three lines a machine can apply. Plus the only claims her marketing may make, each with a source and a date.
What she is betting on this quarter. Two bets, not eight. Each has an audience, one number, and the date she stops.
What happens when a request does not fit the plan
Tuesday morning. Ada’s co-founder writes: we should start a podcast.
A normal skill folder produces a podcast plan by lunchtime. A good one. Nobody asked whether it serves either thing Ada decided to do this quarter.
Here the request hits the briefing step, which has to match it to a bet. It cannot. So it writes this instead of the plan:
Request: start a podcast
Tested against:
Bet 1, own the month-end close question -> no, this is not search demand
Bet 2, outbound on trigger events -> no, reaches nobody with a trigger
Options, in order:
- Drop it.
- Add to the not-doing list, revisit in January.
- Change the quarter. That means killing or shrinking bet 1 or bet 2
to pay for it. Bet 1 has 12 hours a week and 2,000 EUR a month.
A podcast would take most of both.
Option three is last on purpose and it has a price on it. Most requests stop at option one the moment somebody sees that price.
How to run it week to week
The rhythm
- Answer questions, get the two files
One command walks you through it in five steps, one cluster of questions at a time. It reads your repo first, so you are correcting a draft rather than staring at a blank page. You will not finish with perfect answers. You finish with written ones, and gaps marked as gaps.
- It goes through the brief, or it does not happen
Someone asks for something. The brief step matches it to a bet and writes a short brief, or refuses it in writing. This is the only entry point. Nothing gets produced by asking a skill directly.
- It ships, gets spot checked, or waits for you
Internal drafts and research go out on their own. Public work using approved claims is checked one in four. Pricing, named customers, first contact with a named account and paid spend get read every time.
- Read what happened
Each bet gets one of three answers: on track, off track and here is the specific change, or stopped. Plus three numbers about the machine rather than the market: produced against published, how long things sat waiting, and how many claims needed evidence you did not have.
- Decide what was wrong
Stop dates that arrived are stop dates. Then the harder question: is the positioning still true. Claims older than six months, rules that never once fired, customer language that drifted. It proposes the edits. You approve them.
Two things make that rhythm hold rather than drift.
No new numbers. The writing step may only use claims already in the first file, with a source. Anything else never reaches a draft. Most teams do it the other way round and try to catch invented statistics in review. That fails, because a confident invented number reads exactly like a real one at eleven at night on the fourth thing that day.
A cap on production. You promise to approve everything, the stack produces forty pieces a week, you have time for ten, and thirty sit in a folder. Within a month you are rubber-stamping or ignoring it. So you write down the real number, and when the queue passes it the system slows down instead of quietly letting the overflow out.
What is in the plugin
// NINE SKILLS, THREE JOBS
It sends nothing. No email goes out, nothing publishes, no ad money moves. The last action is always a person’s.
It will not make a bad plan work. It will make a bad plan obvious, usually inside three weeks, which is faster than most companies find out any other way.
How to install it
The repository ships a filled-in example, an operator playbook of the positions I take when an answer is vague, and a version of my own tone of voice rules you can keep or replace.
Questions people asked
Can an AI agent replace a CMO?
No. It can take over the production a CMO hands to other people: writing, lists, pages, reports. It cannot do the deciding, which is who you sell to, who you turn away, what gets budget this quarter, and what you stop when the numbers do not move. A stack with nothing written down is a very fast junior team pointed at nothing.
What do you actually have to write down first?
Two files. One says who you sell to, what has to have just happened for them to care, who you turn away, and which claims you are allowed to make. The other says the two or three things you are betting on this quarter, one number each, and the date you stop if it is not working. Half an hour of answering questions gets you a usable first version.
How do you stop an AI from inventing statistics about your company?
Keep a list of claims it is allowed to make, with sources, and let it use nothing else. Catching invented numbers in review does not work, because a confident invented number reads exactly like a real one.
If a human approves every output, does that not just move the bottleneck?
Yes, and that is how most of these projects end. Sort work by risk, spot check the middle tier, and write down how many pieces you can review in a week without skimming. Production slows down when the queue gets longer than that number.
Where the idea came from
In August 2026 Ignacio Prieto posted that he could not afford a CMO, so he built one inside Claude out of seven skill files. It went everywhere. The sharpest thing under it was a comment.
These skills are execution of a marketing plan. Who made the plan?
He shipped a real thing and showed the files, which is more than most people writing about AI marketing in 2026 have managed. My disagreement is narrow: name the first layer after the choice, then make everything else refuse to start until that choice exists.
If you want that written properly for a real company before the automation goes in, that is the work I do.
