Default high think effort causes Opus 4.6 to ignore skills and CLAUDE.md rules
Description
The default think effort setting (high / adaptive thinking) causes Claude Opus 4.6 to ignore user-defined skills and CLAUDE.md rules, severely impacting code generation quality and workflow compliance.
Environment
- Claude Code version: 2.1.34
- Model: claude-opus-4-6
- Platform: Windows 10, Windows Terminal
Problem
With the default high think effort (adaptive thinking), Opus 4.6 frequently:
- Ignores
CLAUDE.mdrules: Custom rules in~/.claude/CLAUDE.mdor project-levelCLAUDE.mdare not followed. The model deprioritizes these instructions when extended thinking is active. - Ignores skill definitions: User-defined skills are bypassed — the model either skips skill invocations or generates output contradicting skill instructions.
- Degrades code output: Results in style violations, skipped workflows, and code that contradicts explicit prohibitions.
Expected Behavior
Opus 4.6 should respect all CLAUDE.md rules and skill definitions regardless of think effort level. Adaptive thinking should enhance reasoning without overriding user-configured instructions.
Actual Behavior
With default high think effort, the model:
- Proceeds with code changes without following
CLAUDE.md-defined workflows - Ignores explicit prohibitions stated in
CLAUDE.md - Skips skill invocations that should be triggered
- Generates responses not conforming to user-specified rules
Steps to Reproduce
- Use Claude Code 2.1.34 with Opus 4.6 on Windows 10
- Configure
CLAUDE.mdwith specific coding rules and workflow requirements - Set up custom skills with clear trigger conditions
- Keep default think effort (high)
- Request code modifications or feature implementations
- Observe that
CLAUDE.mdrules and skill triggers are frequently ignored
Workaround
None known. Manually reminding the model to follow rules partially helps but is unreliable.
Suggested Fix
CLAUDE.md rules and skill definitions should be treated as hard constraints that persist through the adaptive thinking pipeline, not soft suggestions that can be deprioritized during extended reasoning.
10 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
This is specific to Claude Code 2.1.34 + Opus 4.6 combination with the default high think effort setting.
The issue persists and is distinct from earlier reports because:
The pattern is the same (ignoring CLAUDE.md/skills), but this report provides specific version reproducibility information that may help narrow down the regression.
Please keep this open for tracking purposes.
FWIW, this was acknowledged by devs here: https://github.com/anthropics/claude-code/issues/5055#issuecomment-3215519786
...and then they closed the conversation to collaborators. This bug is REAL and should be addressed as a high priority. Opus 4.5 was doing outstanding work for me then same rules, same habits of managing context etc. and 4.6 is making assumptions, not reading the rules, ignoring the rules, doing modifications without asking etc. over and over and over no matter what is in the CLAUDE.md files at any level. 4.5 in my bashrc to the rescue until this model shapes up. I'll take reliable and a little slower EVERY DAY OF THE WEEK over faster and acting like a 14 year old boy. Not that it matters to the company but if 4.6 would have been my first experience with claude I'd have written it off like all the other AI junk.
Confirming this on v2.1.50 + Opus 4.6 (macOS).
I run 13 custom skills with an extensive CLAUDE.md (project-level + global). Today's session:
/clearThis is not a minor annoyance — I rely on skills for production workflows (email triage, project management, ticket handling, content processing). When CLAUDE.md rules are deprioritized, the entire skill system becomes unreliable.
Strongly agree with @mxracer — Opus 4.5 followed instructions reliably. Opus 4.6 is faster but systematically ignores user-defined rules. Please prioritize this.
Possible workaround: would reducing think effort (
/thinkto medium/low) help, since the issue seems tied to adaptive thinking?Addendum to this. Adding instructions to memory also doesn't work because memory itself is a skill and is also ignored making that feature worthless as well.
Closing for now — inactive for too long. Please open a new issue if this is still relevant.
It is inactive because it's not acknowledged or addressed. How about putting the same effort into fixing things that goes into closing github issues?
Also how is it "not planned" to have the model use and follow the guard rails put in place for it? Does this mean it's supposed to be totally autonomous? (that's sarcasm for the bots that may not pick that up)
from
<img width="2017" height="1738" alt="Image" src="https://github.com/user-attachments/assets/08775303-e015-4c84-8fcf-d27ddd4126d7" />
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.