I'm not a programmer. I run a small business with five software projects on one server, and Claude Code builds them. I work mostly at a desktop and by phone one or two days a week.
All I wanted was one central setup, so sessions would stop behaving differently and stop fighting what had already been decided. For weeks that didn't happen.
What went wrong
Every failure got a new check or a new rule. Over five days that meant about 460 commits, and the hooks grew from 28 to 46. The build sat at 4 of its 7 parts.
Our research kept asking how to stop Claude failing, and every answer to that question is another check. Our own notes had already said to freeze first. We never did it.
What we changed
We asked a different question: what do people who finish actually run? We read 20 sources in full, 16 of them from users or forums, then:
- Froze new rules and checks.
- Cut the hooks to five hard stops: destructive commands, schedule changes, bulk data changes, working on the main branch, and editing the guards themselves. Secret keys are refused at commit and merge. A few start-of-chat and warning hooks stay, so 10 hooks in total.
- Fixed the checks that could never pass. A check that can't say yes keeps the work going round forever.
- Switched one project first as the test, then the rest.
- Put the reasons in one central note every chat reads, instead of a copy per project. Without the why, every session argues with what you've already decided.
- Told helpers to make the change first and read afterwards. One helper spent 11 minutes reading on a one-file change and saved nothing. Told to change first, it took about two minutes.
What happened
The cut and all four project switches were merged in about two hours, after weeks of being stuck. The checks on all 7 parts of the build now pass, but an independent review found that overstates it: only 1 part fully meets what I asked for, and 6 are partly there. Some of those checks only prove something exists, and our cut switched off one thing a part asked for. So this is not finished, and I'll report what happens.
What it doesn't fix
Written instructions have no guarantee, and Anthropic's own guide says so. Wording and repeat questions can still go wrong, and the same chat that did this work broke my answer format over and over. The cut also removed warnings that caught some mistakes, and it made mistakes of its own along the way, including removing things I'd never asked it to remove.
What I expect to see
New chats finishing jobs, fewer repeat questions, and no damage getting past the hard stops. I'll update this post with the results, including if it doesn't work.
If you want to try it
List what must never happen and enforce only that. Keep one plain list, one item at a time. Have a person approve the result. After two failed fixes, start clean.
If you're interested, say so in the comments. I can share the structure we put in and a project overview you can paste into your own setup.
This article was written with the help of Claude, from my own project's records. I've checked it and I stand by it.
