A small business running 8 projects on Claude Code for 30 days: rules ignored, projects crossing into each other, and what finally helped

I run a small business. Over the last 30 days Claude Code built and ran about eight projects for me on one Windows office server: an online shop, a CRM, an auction site, a monitoring tool and the server itself. I am not a programmer. I talk to Claude from my phone.

What went wrong

1. My start-up file, my rules and my answer format were written down and loaded in every chat, and still not followed. The answer format came up in 73 of my messages on 23 different days. Doing research before answering came up in 88 messages on 22 days.
2. Chats in one project did another project's work. A server chat ran a shop test on the live shop and stopped publishing for about seven hours, the third time this happened. Another server chat changed my CRM's sign-in. The server project had made 45 working copies of my other projects.
3. Claude asked again for approvals I had already given, and treated questions as approvals.
4. Claude said work was done when it was not. Its own check of 18 of its claims on one morning found 5 true, 12 partly true and 1 false.
5. Most of the work went on rules instead of the business. In 14 days, 46 of 130 merged changes in the shop, 34 of 43 in the CRM, 37 of 47 in the monitoring tool and 24 of 30 in the auction site were about rules, guards and the rules add-on.
6. The GitHub queue jammed. Every change edited the same shared record files, so each merge made every other waiting change clash, and the project's own checks then refused to bring a clashed change up to date.
7. Chats woke up on their own: the desktop app's auto-fix wakes a chat whenever its change clashes, so a chat kept waking, fixing, pushing and clashing again.
8. Anthropic's safety checker refused harmless work, and some things I approved could not be done because of it.

What the research said (189 sources, over half from users and forums)

1. Written rules are requests Claude can skip; only hooks, deny rules and managed settings are enforced (Anthropic's own documents say this).
2. The more rules, the fewer are followed. Anthropic advises under 200 lines.
3. Keep projects apart by the machine, not by a written rule.
4. Build shared pieces once and keep every project linked to them, instead of copying.

How we are fixing it in every project and the whole system

One plan, eight steps, each proved before the next, with real times recorded:

1. Stop projects crossing (done). The server project's 45 copies of other projects were removed. The file every chat loads no longer holds any one project's steps. Work for another project now goes as one complete paste into that project's own chat.
2. Unjam the GitHub queue (in progress). Releases go first. Changes stop editing the shared record files; records are written once after a merge. Checks get spare test databases instead of one shared lock.
3. One central place the machine enforces (done). Claude Code's machine-wide managed settings switch the rules add-on on in every project, keep it updated automatically, and stop a chat writing outside its own project folder. A fresh chat tried to write outside its folder and was refused.
4. One short rule book (done centrally, projects following). The rules every chat reads were cut to 55 lines, sent once by the add-on; the long book is read only when a job needs it. Each project drops its own copy of the rules.
5. Cowork and the phone (done). The same short rules went onto the account's Cowork add-on, and the account-wide instructions now hold the current answer rules; the old ones that contradicted them were removed.
6. Build once, use everywhere (in progress). One home for shared pieces; nothing is shared until it has worked in two projects; an hourly scorecard checks every project against the central setup and shows which ones have not taken it yet.
7. Tell people (this post).
8. The habits that keep it fixed: one chat writing to a project at a time; a chat that goes wrong is replaced by a fresh one with one full brief, not patched; approvals recorded where the machine reads them; the plan kept as a tick-list a chat can only tick; tests not changed to make work pass; scheduled jobs off the exact hour.

The first five steps took about two and a half hours of Claude's own time, against an estimate of eight to nine hours.

What we would like Anthropic to fix

1. A way to make the answer format and start-up rules hold in chat and Cowork, not only in Claude Code hooks.
2. Auto-fix should not wake a chat in a loop when its change keeps clashing.
3. The safety checker should honour an approval the owner has already given in the same chat.
4. A way to open a chat in a chosen project from the phone.

Earlier reports from the same setup: anthropics/claude-code issues 107180 to 107184 (4 Nov 2026).

Also reported to Anthropic on GitHub: https://github.com/anthropics/claude-code/issues/108568