I rewrote my Claude Code session rules 41 times in six weeks. Here’s what kept breaking
retained text
I run a bunch of Claude / Claude Code sessions in parallel across a few projects. Every time one burned me, I wrote a rule. Six weeks, 41 rewrites. Yeah. The ones I think other people will recognize: Two sessions, one name. A second session opened under a chat ID that was still live — and passed every start check I had. The only thing that caught it was a git status. Now nothing writes until it checks for a twin. 123 passing tests. Three audits. Still broken. The output shape was one the downstream engine flat out refused, and nothing caught it bc everything checking it was built by the thing being checked. The builder's own tests don't close a phase anymore — a fresh session or a live run does. The audit that didn't audit. The reviewer model couldn't read the files in its sandbox, and the skipped check came back looking clean. So now: not run in full = not a pass. Partial runs say PARTIAL on the first line. It grepped the manual. A session searched the instruction file for the one step it wanted instead of loading the whole thing. (Reading one page and calling it read.) Invoke first, then act. Zombie opener. A session re-ran the startup routine for a chat that had already ended, and started re-issuing work. Approval in the wrong box. I approved from my phone and my typed "yes" landed in the wrong session, not the session that was waiting on it. Approvals are a card now, never free text. The pattern usually isn't bad code, but something reported as done that nobody verified. I've got a few dozen more of these. I’ve been thinking about writing them up as a short playbook (incident, rule, drill). Worth it? And which of these have you hit?
document fetched 2026-11-14 00:44 · canonical https://reddit.com/r/ClaudeCode/comments/1xvc77s/i_rewrote_my_claude_code_session_rules_41_times