Per-project quality variance: One project consistently fails while others work fine

Hi everyone,

I've filed several issues here before about rule compliance, hook circumvention, and session memory (#29795, #29692, #29691, #29689, #29709). This is a different observation.

The pattern:

I run Claude Code across multiple projects on the same machine, same model (Opus 4.6), same setup. Most projects work reasonably well. One project has been failing systematically for 14+ sessions. Same types of mistakes, over and over, despite:

• Extensive CLAUDE.md rules
• 10+ feedback memories from prior corrections
• Documented failure catalogs from previous sessions
• Technical enforcement hooks

The failures aren't exotic. They're basic instruction non-compliance: exposing information that rules say not to expose, skipping steps that are marked as mandatory, recognizing a rule and then violating it in the very next action. The AI documents its own failure, saves a rule about it, and then repeats it immediately.

Other projects with comparable or greater complexity don't have this problem — at least not at this scale.

What I can't explain:

The project isn't more complex than the others. The rules aren't more demanding. The instruction surface is larger because of accumulated failure documentation — but other projects have similar volumes without this degradation. It feels like this specific project's context has entered some kind of failure loop that more instructions can't fix.

My questions:

1. Does anyone else see consistent quality differences between projects that aren't explained by complexity?
2. Has anyone observed a project where accumulated corrections and failure documentation made things worse instead of better — as if the negative context itself degrades performance?

Any data points appreciated.