Skip to content
AgentFix

Adoption LessonsBuild3 min read

We built this site by talking to Claude. Even the merge.

Jason Lawrence

We rebuilt this site in September by talking to Claude. The positioning and the page copy were written and approved by people. Claude wrote the code from that copy, and nearly every change since. When a change was ready, I’d type “looks good, merge it”, and Claude merged it.

That sounds like the moment a leader should start worrying. It wasn’t, and the reason isn’t trust in the model.

The number: 57 pull requests, each approved by a person

A pull request is a proposed change with a record attached: exactly what changes, why, and a preview of the result. Nothing reaches the site without one.

By early October, Claude had co-written 110 of the 111 changes on the branch that becomes the live site. And 57 pull requests had been merged, every one approved by me. Some I merged with a click. Some I approved by telling Claude to merge them. The history can’t tell the two apart, and it doesn’t need to. Either way, a person looked first and the record says what they approved.

What made “merge it” safe

Every change takes the same path. Claude makes the change on its own copy of the site. That copy gets its own preview address. I look at the preview. The pull request records what is changing. Only then is there anything to merge.

The guardrail is in the tooling, not in a document. Claude is blocked from changing the live branches directly, from deleting them, and from changing the rules that protect them. If it tries, the setup stops it and tells it to open a pull request instead. Claude can merge when asked, but only something that came through that path.

Production is one more step away. Merged work goes to a shared preview first. The live site takes it in a separate release, which I approve on its own.

So “merge it” in a chat is an approval of something specific: a change I’ve seen, in a preview I’ve checked, with a record that outlives the chat.

What went wrong

The path came after the failure, not before it. In the first week, work piled up on parallel copies of the site. By 26 September four previews showed four different sites. One had the new favicon, another the redirect fixes, and none had everything. Conversation made changes fast. It didn’t keep them in one place.

The fix was one shared branch that everything goes through, and one preview to show people. The rule now is that nothing reaches it without a pull request.

The other half is the Git lesson: the hard part of building with Claude was never Claude.

What we would do now

Set up the path before the first change, not after the fourth preview. Decide where approvals are recorded, and make sure the tooling, not a policy line, stops anything skipping it.

Then let the approval be as easy as a sentence. The sentence isn’t the risk. A sentence with no record behind it is.

One thing to try this week

If anyone on your team builds with Claude, ask them where the record of each approval lives. If the answer is “in the chat”, that’s the gap to close first.

“A human owns every output” is one of the five rules the Adoption Diagnostic checks your AI policy against. About the Adoption Diagnostic

More lessons

Get new lessons by email

Short lessons from a real Claude rollout, each backed by a number.

A new lesson every Tuesday. Unsubscribe any time. Handled as set out in our privacy policy.