Hey, thanks for the detailed report.
This is a known limitation of how rules work right now. Rules are passed to the model as instructions in the system prompt, but models are non-deterministic, so they might ignore constraints, especially negative ones like “don’t do X”. The team is aware of the issue.
The most practical workaround is to rewrite rules in a positive way and explain why. Models follow positive instructions with context much better than plain bans.
Instead of:
“Don’t write documentation, just explain in chat”
Try something like:
“Your role is to explain solutions directly in chat conversation. This is more helpful because the user can ask follow-up questions immediately. When you need to share code changes, propose edits to source files. Never create or modify .md documentation files, always explain findings and solutions as chat messages.”
A couple more tips:
- Shorter chats are more reliable, in long sessions rules get “forgotten” more often
- Concrete step-by-step instructions work better than abstract bans
These threads have similar cases and more detailed guidance from the team:
Let me know if rewriting the rules doesn’t improve things.