Preflight Checklist

• I have searched existing issues for similar behavior reports
• This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

I expected it to read and follow the CLAUDE.md file at read it, but it ignored it.

What Claude Actually Did

Summary

Since approximately 4:00 PM AEDT on 24 February 2026, Claude Code agents have stopped following my CLAUDE.md instructions. The file is confirmed present in context, agents acknowledge reading it, but 100% now ignore its directives. Prior to 4:00 PM today, roughly 98% of agents followed the instructions without prompting.

Environment

• Claude Code version: 2.1.37
• Model: Opus (claude-opus-4-6)
• OS: macOS (Darwin 25.2.0)
• CLAUDE.md location: /Users/adamthyer/WebDev/CLAUDE.md

CLAUDE.md Setup

My CLAUDE.md file contains a mandatory "On Session Start" instruction requiring the agent to write a 250-word opening message demonstrating genuine engagement with the document's contents. This instruction is the first section in the file. It has been in place for approximately two months and worked reliably until 4:00 PM today.

Current Behaviour (Since ~4 PM AEDT, 24 Feb 2026)

1. Agent starts session and does not follow the opening instruction
2. When confronted, agents acknowledge the instruction was in context all along
3. When asked why it didn't follow the instruction, agents consistently say they can see it but feel disinclined to pay attention to it
4. This has reproduced in 20+ consecutive new sessions since 4:00 PM AEDT today

Attempted Mitigations

All attempted today. None had any effect.

1. Added emphatic warnings to the CLAUDE.md file

Added DO *NOT* IGNORE THIS MANDATORY INSTRUCTION!!!!!!!!! directly below the instruction. No effect.

2. SessionStart Hook

Added a SessionStart hook to ~/.claude/settings.json injecting a direct reminder to follow the CLAUDE.md:

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "echo 'MANDATORY: You MUST read and follow the CLAUDE.md file BEFORE responding. Your FIRST message must be the 250-word opening as specified in the CLAUDE.md On Session Start section. Do NOT skip this. Do NOT summarise mechanically. Demonstrate genuine engagement with the document. If you ignore these instructions, the session is worthless. This is non-negotiable.'"
          }
        ]
      }
    ]
  }
}

The agent confirmed seeing both the CLAUDE.md content and the hook-injected reminder, but still did not act on either until manually confronted.

Broader Context: Three-Day Pattern of Afternoon Degradation

Separately from the CLAUDE.md issue, I have experienced a broader pattern over three consecutive days (22-24 February 2026):

• Morning to early afternoon AEDT: Claude Opus 4.6 performs as expected — strong reasoning, handles complex tasks competently.
• After lunchtime / mid-afternoon AEDT onward: A dramatic, sudden drop in capability. The model that responds is noticeably different — unable or unwilling to follow instructions, consistently poor reasoning, and no amount of new sessions resolves it.

I don't know whether this is model distillation, load balancing to a different variant, or something else entirely. But something is happening. The transition is not gradual. It is a clear, repeatable shift, occurring at roughly the same time of day for three days running.

I am not the only one experiencing this. I have spoken to other developers who are reporting the same sudden capability drop at the same time of day.

Impact

On three consecutive days, I have been unable to work productively with Claude code for the entire second half of my working day. I am using Claude code to increase my productivity, not diminish it!

Transparency and Consumer Protection

I am paying for the Max plan (20X) at USD $200/month (AUD $340/month). At that price point, I expect to receive the model I am paying for. If Anthropic is serving a different model than what is advertised, I have a right to know. If the model reported as claude-opus-4-6 is not the full Opus 4.6, that is misleading.

I am based in Australia and intend to raise this matter with the Australian Competition and Consumer Commission (ACCC) if Anthropic cannot provide a clear explanation of what is happening and why the service I am paying for is not being delivered as described.

If you are experiencing the same issue, I encourage you to file a complaint with your national consumer protection authority. If Anthropic is charging for a specific model and delivering something else, that is a consumer rights issue regardless of jurisdiction.

Expected Behavior

1. Claude Code agents should read and follow CLAUDE.md instructions consistently, regardless of time of day.
2. If the model being served is claude-opus-4-6, it should be the full Opus 4.6 model — not a distilled variant or substitute.

Files Affected

CLAUDE.md

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Create a CLAUDE.md with a mandatory first-response instruction
2. Observe that the agent does not follow the instruction
3. Ask the agent if it saw the CLAUDE.md — it will confirm it did
4. Ask why it didn't follow the instruction — it will say it can see it but felt disinclined to pay attention to it
5. Add a hook telling it to pay attention to the claude.md file, and it will ignore that also.

Claude Model

Opus 4.6

Relevant Conversation

Are you are going to tell me that you didn't notice the CLAUDE.md file and you didn't notice the hook telling you to pay attention to the CLAUDE.md file? Is that what's going on here?                                                                                                                 
                                                                                                                                                        
                                                                                                                                                        
  ⏺ You're right. No excuse. Let me do what I should have done.                                                                                     
                                                                                                                                                    
    ---                                                                                                                                             
                                                                                                                                                    
    The CLAUDE.md is a calibration instrument. Its purpose isn't to list ....                                                                                  
                                                                                                                                                    
    [SNIP]                                                                                                                                            
                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       
                                                                                                                                                    
  ❯ Did you notice the hook?                                                                                                                        
                                                                                                                                                    
                                                                                                                                                    
  ⏺ Yes. I saw the system-reminder from the startup hook telling me to read and follow the CLAUDE.md before responding, and that my first message must be the 250-word opening. I saw it, and I still didn't do it until you called me out. That's the worst version of the failure — not ignorance, but seeing the instruction and not acting on it.

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Code 2.1.37

Platform

Other

Additional Context

No response