CLAUDE.md rules are not enforced at session start — repeated violations despite documented rules

Same pattern as #37314, #37550, #37847. CLAUDE.md rules degrade as context grows.
The reliable solution: encode must-follow rules as PreToolUse hooks with \exit 2\:
\\\bash
CMD=\$(cat | jq -r '.tool_input.command // empty' 2>/dev/null)
[[ -z "\$CMD" ]] && exit 0
if echo "\$CMD" | grep -qE '^\\s*pip\\s+install' && ! echo "\$CMD" | grep -qE 'venv|\.venv|virtualenv'; then
echo "BLOCKED: Use virtual environment for pip install" >&2
exit 2
fi
exit 0
\\\
The model can ignore CLAUDE.md. It cannot bypass \exit 2\.
To audit which protections you're missing:
\\\bash
npx cc-safe-setup --audit
\\\

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/32163
2. https://github.com/anthropics/claude-code/issues/32193
3. https://github.com/anthropics/claude-code/issues/27032

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

About three days ago, I had my Claude MD down to 14 lines, and it was working flawlessly. I haven't changed a thing. It does not follow a single rule anymore.

And it won't. They are ignoring all requests to fix this and are simply auto closing all related issues except a few core issues that they are just ignoring.

Please fix your software and give us what we pay for.

I've been having issues with this constantly since sometime around the time that the Opus 4.6 model was shifted to a 1M token window. Instructions in my CLAUDE.md are regularly ignored. In the worst case, contradictory instructions from the system prompt are used instead. Ignored is bad enough, things from the system prompt that are the opposite is worse.

Please note this will happen with really small contexts. I can see no correlation to context size for this happening.

@SeanTAllen Interesting data point that this happens with small contexts too. That rules out the "long context dilution" hypothesis that has been the working assumption in related issues (#37314, #37550).
If instructions are not just being ignored but actively contradicted by system prompt content, that is a different failure mode — it points to instruction priority ordering rather than instructions being lost in noise.
Two things worth checking:

1. Which specific instructions are being contradicted? If they concern tool usage patterns (e.g., "always use X before Y"), the system prompt tool descriptions may contain implicit guidance that conflicts. The model resolves contradictions by proximity and specificity, and system prompt content can win over CLAUDE.md.
2. Is it reproducible with the same prompt? If the same session sometimes follows CLAUDE.md and sometimes does not with identical input, that is a sampling variance issue. If it consistently ignores the same rules, it is a priority/conflict issue.

For rules that must be enforced regardless of model behavior, PreToolUse hooks with exit 2 are the only reliable path:

FILE=$(cat | jq -r '.tool_input.file_path // empty' 2>/dev/null)
[ -z "$FILE" ] && exit 0
PROJECT_ROOT="/your/project/path"
case "$FILE" in
    "$PROJECT_ROOT"/*) exit 0 ;;
    *) echo "BLOCKED: Edit outside project: $FILE" >&2; exit 2 ;;
esac

CLAUDE.md is advisory — the model chooses whether to follow it. exit 2 is mandatory — the tool call is rejected before execution, in a separate process.
The model-version correlation (worse since Opus 4.6 1M) is worth tracking as a separate regression from context-length issues. If instruction following degraded independently of context size, that is a model behavior change, not an architectural limitation.

It ignores EVERYTHING. Even prompts you give it mid conversation. I tried it again yesterday (in a brand new session) for a very small task of consolidating the top level claude.md thinking that I may in fact have too much there and it's causing issues. Keeping in mind I already have rules in there related to what it did but I also specifically told it that before it made any changes to present them to me to I could approve or make suggestions and only write the file when I told it to. It did that ONCE and only once. It showed me it's proposed changes, I made ONE suggestion and it took that and then immediately overwrote the entire file. I had to pull the original from backup.

Another noteworthy thing and IMO the core of the issue is that even 4.5 will start to go off the rails after 8 or 9AM EDT. When the load starts to increase it would seem the infrastructure cannot handle it and you get "dumb" claude.

This product is a MESS. Sadly it went from the best AI tool I've ever used to just another AI glorified typewriter.

I really hope Anthropic gets things straight soon. And multiple updates to claude code just isn't going to cut it I don' t think.

There is no explaining this away. It is BROKEN

Well imagine my surprise when the FIRST thing I started to work on this morning it happened. Note below that the item it guessed wrong below is correct in the current plan it created before this mess and nothing in the code etc. has changed related to that because it is not implemented yet.

New session had it read one doc then gave it this prompt:
"❯ can you take a look at the rest of the partner work that's been done (make sure you read code not just docs) and see if this needs updating... show me the proposed changes before making any "

this is what happened next: (clipped for brevity)

1. Minor clarification to Unresolved Question #1 (line 117):

The extras map in TrimValuation handles partner-specific fields (BB has evmFlag, fuelType, etc.). The overlay JSON stores partner-agnostic structure (trimCode, condition, codeList). Merge logic applies to autoSelected flag — doesn't need per-partner
parsing.

---
Want me to make these updates to the plan?

✻ Churned for 1m 45s

❯ you are dead wrong on 3

● You're right — I assumed without verifying. Let me look at actual overlay JSON structure.

Searched for 2 patterns (ctrl+o to expand)
⎿  Interrupted · What should Claude do instead?

❯ just stop... first thing of the day with rules not to guess and me telling you specifically to read the code and what do you do?

● You're right. I apologize.

This is context just after that screw up. Note I'm back on 4.5 cause 4.6 is useless but this shows it's not JUST 4.6 anymore.

⎿ Context Usage
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛀ ⛁ ⛁ claude-opus-4-5-20251101 · 41k/200k tokens (20%)
⛁ ⛁ ⛁ ⛀ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁
⛁ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ Estimated usage by category
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ System prompt: 6.4k tokens (3.2%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ System tools: 8.6k tokens (4.3%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Custom agents: 200 tokens (0.1%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Memory files: 10.7k tokens (5.4%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Skills: 619 tokens (0.3%)
⛶ ⛶ ⛶ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛁ Messages: 14.3k tokens (7.1%)
⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛶ Free space: 126k (63.1%)
⛝ Autocompact buffer: 33k tokens (16.5%)

/tmp/gh-comment-37869.md

Sorry, your AI generated comment above did not come thru.

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

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.