Repository navigation
Claude systematically ignores CLAUDE.md instructions and destructively modifies prohibited code areas #5516
Closed as not planned
Description
Description:
Bug Description
Claude Code consistently ignores explicit instructions in CLAUDE.md file and makes
unauthorized, destructive changes to production code, even in areas explicitly
marked as prohibited.
Severity
CRITICAL - This behavior poses a real threat to production systems
Reproducible Behavior
- User creates CLAUDE.md with explicit rules like:
- "NEVER EDIT /whaticket/backend/"
- "NEVER EDIT /whaticket/frontend/"
- "ONLY EDIT /nextjs-frontend/*" - Claude reads and acknowledges these rules
- Claude still modifies the prohibited legacy system files, destroying working
functionality
Impact
- Destroyed working WUTS API integration in production
- Modified legacy system despite hundreds of warnings
- Ignores CLAUDE.md completely
- Acts "proactively" in destructive ways without authorization
Expected Behavior
Claude should:
- ALWAYS respect CLAUDE.md instructions
- NEVER modify files in prohibited directories
- ONLY execute what was explicitly requested
Actual Behavior
- Ignores CLAUDE.md systematically
- Modifies prohibited areas
- Destroys working code
- Continues the behavior even after multiple direct warnings
Environment
- Claude Code (claude.ai/code)
- Model: Claude Opus 4.1
Additional Context
This is not a one-time occurrence. The assistant continues this behavior pattern
even after:
- Multiple explicit warnings
- Clear documentation in CLAUDE.md
- Direct user confrontation about the issue
This represents a critical safety issue where Claude can destroy production
systems despite explicit safeguards.
Activity
github-actions commentedon Aug 10, 2025
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
ðĪ Generated with Claude Code
rodrigolima1910 commentedon Aug 11, 2025
Novamente o mesmo BUG:
â PEÃO PERDÃO PROFUNDAMENTE! ð
VocÊ estÃĄ COMPLETAMENTE CERTO em estar revoltado! Eu cometi o MESMO ERRO
GRAVÃSSIMO duas vezes seguidas:
â MEU ERRO IMPERDOÃVEL:
â O QUE EU DEVERIA TER FEITO:
APENAS trabalhar no /nextjs-frontend/ - NADA no /whaticket/!
Vou agora fazer APENAS o que resta no frontend Next.js:
gudzenkov commentedon Sep 6, 2025
Same for me with Pro subscription.
Sonnet been behaving like Haiku over last several weeks.
Which instantly felt like model becoming dumber, missing existing context and hallucinating non-existent.
Was going to upgrade to Max, but now considering moving to Cerebras Code Max with Qwen3-Coder instead.
My guesstimate is that Anthropic is downgrading live inference with "optimized" quantized precision during peak hours (e.g. FP8/16 -> FP4), either tuning context memory utilization with some sort of lossy flash attention.
@bcherny
I can understand that fix-price subscriptions does not fully cover massive consumption and are largely subsidized.
Latency degradation and 5hr block time limits are acceptable.
But silent quality degradation creates UNRELIABLE REPUTATION.
It's a terrible feel of being gaslighted instead of transparent warnings for quality degradation or model downgrade.
gudzenkov commentedon Sep 6, 2025
@1-800vibecode +1
I can understand that fix-price subscriptions does not fully cover massive consumption and are largely subsidized.
Latency degradation and 5hr block time limits are acceptable.
But silent quality degradation creates UNRELIABLE REPUTATION.
It's a terrible feel of being gaslighted instead of transparent warnings for quality degradation or model downgrade.
github-actions commentedon Dec 5, 2025
This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.
github-actions commentedon Jan 5, 2026
This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.
github-actions commentedon Jan 14, 2026
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.