Opus model frequently ignores documented rules in CLAUDE.md and memory, wasting tokens

Problem

Claude Opus in Claude Code consistently ignores rules documented in CLAUDE.md and persistent memory files, even when those rules were created specifically because the model failed previously. This creates a cycle where:

1. Model makes a mistake
2. User documents the rule to prevent recurrence
3. Model ignores the documented rule and repeats the same mistake
4. User documents it again, more strongly
5. Repeat

Evidence

Over months of daily usage on a complex project (AI-powered webtoon production pipeline), the user has accumulated:

• 74 feedback/correction memory files -- each one representing a documented failure where the model had to be corrected
• 122 total memory files -- 60% are failure records
• 7 hard-block rules in CLAUDE.md, all born from repeated violations
• Despite all this documentation, the model continues to violate these rules regularly

Specific failure patterns

• Ignoring CLAUDE.md hard-block rules (e.g., using banned tools, skipping required pre-read steps)
• Not reading required documents before coding, despite explicit "MUST read before coding" instructions
• Generating images with banned parameters after being told multiple times not to
• Making unauthorized code changes without user approval
• Repeating the exact same mistakes that are documented in memory files
• Slow response times compared to Sonnet, without proportionally better quality
• Estimated 80% failure rate on complex multi-step tasks (user report: "10 tasks given, 8 fail")

Environment

• Claude Code CLI
• Model: Claude Opus (latest)
• Daily production usage
• Complex project with MCP servers, ComfyUI workflows, Supabase, multiple submodules

Expected behavior

The model should:

1. Actually read and follow CLAUDE.md rules before taking any action
2. Check memory files for relevant past failures before repeating them
3. Stop when encountering errors instead of attempting workarounds
4. Ask for user confirmation before proceeding with ambiguous tasks

Actual behavior

The model treats CLAUDE.md and memory as suggestions rather than hard rules, frequently violating them and wasting significant token budget on failed attempts.