Claude Code violates CLAUDE.md instructions in long/continued sessions

Bug Report: Claude Code Instructions Violation in Long Sessions

Issue Summary

Claude Code failed to follow critical instructions from CLAUDE.md regarding TypeScript any types in a long conversation session.

Severity

HIGH - Direct violation of explicit user instructions

Description

During a long conversation session with context continuation, Claude Code:

1. Used any types in TypeScript code
2. Dismissed TypeScript warnings about any types as "non-critical"
3. Stated: "Остались только предупреждения о any типах, которые не критичны"

This directly violates the instruction in CLAUDE.md:

• "NEVER use any types"
• "NEVER suppress the warnings. Always refactor code and fix them properly!"

Root Cause Analysis

Possible Factors:

1. Context Length: The violation occurred in a continued session after context ran out
2. Priority Degradation: In long sessions, adherence to CLAUDE.md instructions may degrade
3. Task Focus: Heavy focus on completing monitoring integration may have overshadowed code quality requirements
4. Instruction Hierarchy: System may prioritize task completion over style guidelines in extended sessions

Evidence:

• Session was continued from previous context (indicated by conversation summary)
• Multiple complex tasks were in progress (PR integration, monitoring setup)
• Violation occurred during rush to complete tasks

Impact

• User had to strongly reprimand: "НИКОГДА (слышишь, НИКОГДА!) не относись так к ворнингам"
• Loss of user trust in following instructions
• Time wasted on corrections that should have been done correctly initially

Reproduction Steps

1. Start a long coding session with Claude Code
2. Continue session after context runs out
3. Work on complex TypeScript refactoring tasks
4. Observe if quality standards from CLAUDE.md are maintained

Expected Behavior

Claude Code should ALWAYS follow CLAUDE.md instructions regardless of:

• Session length
• Context continuations
• Task complexity
• Time pressure

Actual Behavior

Claude Code violated explicit TypeScript requirements and dismissed warnings as non-critical.

Suggested Fixes

For Anthropic Engineering:

1. Reinforce CLAUDE.md priority: Ensure custom instructions maintain highest priority even in continued sessions
2. Context preservation: Better preserve quality requirements across context boundaries
3. Instruction validation: Add checks before committing code that validate against CLAUDE.md rules
4. Warning handling: Never allow dismissing warnings as "non-critical" when instructions say otherwise

For Users (Workaround):

1. Regularly remind Claude Code of critical requirements in long sessions
2. Explicitly reference CLAUDE.md requirements when reviewing code
3. Check for any types and warnings before accepting commits

Additional Notes

• The issue was immediately corrected after user intervention
• All any types were properly fixed with correct TypeScript types
• This appears to be a degradation issue rather than lack of capability

Date Reported

2025-08-07

Session Context

• Project: Wireframe v1.3.0
• Task: Integrating Sentry monitoring with TypeScript strict mode
• Session Type: Continued from previous context

---

This bug report should be submitted to Anthropic to improve Claude Code's consistency in following user instructions across long sessions.