S2E06: I Handed Over a Website With No CMS. On Purpose.

The most AI-native thing I've shipped this year is a website handover.

The website itself was the easy part. I built it for a live-commerce client, launched it on his domain, sent over a deck, a document, and a guide. Stepped back. Standard agency theatre so far.

Except the site has no CMS. No admin panel. No WordPress login. No plugin garden to water every weekend. When he wants to publish a post, he opens Claude Code and talks.

I know how that sounds. I removed the thing every website is supposed to have and I'm calling it a feature. Stay with me.

How he publishes now

Two agents live in the repo.

The first one reads his actual operations data. Live session results, shop numbers, what happened on the ground this week. Then it proposes stories. Not "content ideas". Things that already happened, with his own numbers attached. The internet does not need another 5-tips post, and he definitely doesn't need to write one.

The second agent drafts and publishes. It cannot publish anything without his yes. That's the entire editorial process: one man saying yes or no.

Here's my favourite part. Every yes and every no gets logged to a taste file, with his reason, word for word. The proposing agent reads that file before pitching again. Every rejection is a training signal.

I basically built him an intern that takes rejection well and remembers why.

The publishing loop: rules stay frozen; taste compounds.

Rules don't learn. Taste does.

The taste file sits next to a rules file. They never mix.

Rules are frozen: anonymise by default, every number comes from a real export, always draft first. These never change, no matter how many drafts get approved.

Taste learns: which stories he picks, what he kills at preview, which words he crosses out.

Why keep them apart? Because a rule that can learn is just a suggestion with good posture. Enough charming drafts and it starts to bend. So the rules don't get to learn. They sit there being rules, which is the whole job.

The part that rearranged my head

I've been saying "AI-native" in client meetings for months. This isn't my first system like this. But it's my best example, because a company blog is the most boring asset a business owns. And even there, the workflow itself changed shape.

Knowledge sits in the system, not the person. The guide, both agents, the taste file. All in the repo. If he hires a marketing exec next year, they clone the repo and the whole operation comes with it. Nothing walked out in my head when I stepped away. Handover used to mean writing documentation nobody reads. This time the system is the documentation.

Approval is the publish button. His whole job in the pipeline is yes or no. Sounds like a demotion until you watch it. Drafts are cheap now, the agent will happily make forty. Knowing which one sounds like him in front of buyers is the scarce part. The system runs on his judgement, not his typing.

The 90 days

I gave the system a 90-day monitoring window before we decide what ongoing support looks like. Standard stabilisation clause. Very professional. I even wrote "observe, tune, document" like a person with a methodology.

Halfway through the handover doc I noticed what the window actually is: a training period. Every week it runs, the taste file grows. By day 90 that file is a distilled profile of how one operator decides what's worth saying in public.

That file might end up worth more than the website, which is a strange thing to admit after building the website.

The thing I've been selling wrong

In client meetings I sell AI as speed. Faster drafts, faster launches, faster reports. And sure, it is fast.

But two better things changed here: who holds the initiative, and where the human sits.

The initiative belongs to the machine now. It doesn't wait for him to feel like blogging. It watches his operations, harvests the stories worth telling, and shows up with drafts. Publishing no longer costs him his evenings. It costs him a decision.

And that decision sits at the one spot that maximises his judgement. A yes publishes the post. A no teaches the system what to stop pitching. One decision, two effects: it ships and it trains. Thirty seconds of him buys more there than an hour of him anywhere else in the pipeline.

That's what AI-native looks like: machines holding the initiative, people parked where their judgement pays the most.

One decision does two jobs: it ships the post and trains the system.

One person in the loop works beautifully. Next I want to watch what happens when he adds a second.