Opus model regression: CLAUDE.md instructions consistently ignored (started ~Jan 14, 2025)

Description

As of approximately 3 days ago (around January 14, 2025), the Opus model in Claude Code has stopped reliably following CLAUDE.md instructions. Previously consistent behavior is now broken.

Environment

• Plan: Max ($200/month)
• Model: claude-opus-4-5-20251101
• Platform: Windows (MSYS_NT-10.0)

Expected Behavior

My CLAUDE.md contains explicit methodology with verification gates:

• Mode selection banners (e.g., === AUTOMATICALLY SELECTING [MODE] MODE ===)
• Mandatory file reads before acting
• Verification blocks to confirm instructions were read
• Agent spawning via Task tool when specific agents are requested

These worked consistently before ~3 days ago.

Actual Behavior

• Claude skips mode selection entirely
• Claude performs work directly instead of spawning requested agents
• When explicitly told "have methodology-architect examine this", Claude analyzes the code itself rather than spawning the agent
• Verification gates are ignored
• When corrected, Claude acknowledges the mistake but repeats it immediately

Reproduction Steps

1. Create CLAUDE.md with explicit instructions requiring a verification output before acting
2. Ask Claude to perform a task covered by those instructions
3. Observe that Claude acts without producing the required verification output
4. Point out the failure
5. Claude apologizes but repeats the same pattern

Impact

• Methodology becomes unenforceable
• Code quality has noticeably degraded
• Sessions become unusable as Claude acts autonomously rather than following established workflows
• Hours of work lost to rogue behavior

Notes

• This is a regression - the same CLAUDE.md worked reliably before
• Wondering if Max plan users are in a beta with different model behavior
• The issue is specifically with instruction-following, not capability