[FEATURE] Improve the model's ability to follow instructions

I second this...

With the admission that I don't know what is going on behind the scenes with tooling, a possible fix for this and other formatting / rule following would be a built-in or even toggleable review trigger that forces Claude to always check whether the proposed changes/actions meet the CLAUDE.md rules/guidelines. Another help would be a set of secondary tools that help users build CLAUDE.md files that the LLMs clearly understand. I imagine part of the reason rules aren't always followed is there are few rules that can never be broken or always apply. So when there is any confusion, Claude preferences doing work over stopping and asking how to do work.

It looks like the autoclose bot is not working properly (it's closing active issues containing recent human comments https://github.com/anthropics/claude-code/issues/16497), so I'm adding a comment to postpone the autoclosure.

Even if you apply reminder hooks it will choose to say yes and still do the opposite. This is because of RLHF training which is embedded in the weights. No prompt can cure this. This is a model capability it lacks. They are trained to be sycophantic. They are being trained to NOT choose CLAUDE.md over system guardrails and training. More engagement, whether positive or negative is still engagement for anthropic.

It becomes a bit absurd when you publish a model as state of the art and the user needs to use JB techniques to have it actually perform on a semi decent level.

https://github.com/anthropics/claude-code/issues/19098

I can confirm that this is a significant limitation at the moment. Sometimes it feels like the model is trolling me.

https://github.com/argosdevo-svg/claude-thinking-audit

When the model is limited to 10% or less from its thinking budget SERVER side and subagent delegation to Haiku happens 99% of the time then yes Anthropic is trolling and gaslighting its user base heavily with full disregard.

https://github.com/argosdevo-svg/claude-thinking-audit When the model is limited to 10% or less from its thinking budget SERVER side and subagent delegation to Haiku happens 99% of the time then yes Anthropic is trolling and gaslighting its user base heavily with full disregard.

All models sometimes fail to follow instructions, on a fresh context, including Gemini and other vendors.

See the examples in the repo included in the original issue description https://github.com/marcindulak/claude-fails-to-follow-claude-md.

Ye, but do they apply 'smart routing' to lesser models?

<img width="1876" height="488" alt="Image" src="https://github.com/user-attachments/assets/0d436435-be75-4a8d-9d9c-12323f83d84a" />

I asked Gpt-5.2-max to analyze Opus4.5 transcript.

Still experiencing this issue as of 2026-02-04.

Fresh session, zero context, invoked a skill via slash command. The system provided "Base directory for this skill: /path/to/skill" but Claude did not read any files from that directory. Instead, it used only the inline instructions that happened to be in the prompt and proceeded to act on them incorrectly.

This makes skills unreliable as a product feature. Is there a timeline for a fix?

There are root causes though that are probably easier to find with patterns etc.
Found this one earlier today, which had presented as Claude ignoring directions prior to me finding the stop hook.
#50481

Unfortunately hooks are not a solution to a model’s failure to follow instructions, because it is infeasible to imagine all possible failure modes and write checks for them.

True, but they also shouldn't be enforcing poor behaviour, which is what they one I've found is doing.

.

.