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

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude Code fails to follow instructions from CLAUDE.md.

See examples of existing issues:
https://github.com/anthropics/claude-code/issues/528 https://github.com/anthropics/claude-code/issues/668#issuecomment-3036602220 https://github.com/anthropics/claude-code/issues/742 https://github.com/anthropics/claude-code/issues/2142 https://github.com/anthropics/claude-code/issues/2544 https://github.com/anthropics/claude-code/issues/2670#issuecomment-3036644347 https://github.com/anthropics/claude-code/issues/2901 https://github.com/anthropics/claude-code/issues/3377 https://github.com/anthropics/claude-code/issues/3382#issuecomment-3105883327 https://github.com/anthropics/claude-code/issues/5055#issuecomment-3215519786 https://github.com/anthropics/claude-code/issues/5502 https://github.com/anthropics/claude-code/issues/6059 https://github.com/anthropics/claude-code/issues/7571 https://github.com/anthropics/claude-code/issues/7777 https://github.com/anthropics/claude-code/issues/8043 https://github.com/anthropics/claude-code/issues/8059 https://github.com/anthropics/claude-code/issues/8738 https://github.com/anthropics/claude-code/issues/12503 https://github.com/anthropics/claude-code/issues/12832 https://github.com/anthropics/claude-code/issues/14987 https://github.com/anthropics/claude-code/issues/16073 https://github.com/anthropics/claude-code/issues/16506 https://github.com/anthropics/claude-code/issues/16527 https://github.com/anthropics/claude-code/issues/17228 https://github.com/anthropics/claude-code/issues/17241 https://github.com/anthropics/claude-code/issues/17530 https://github.com/anthropics/claude-code/issues/17616 https://github.com/anthropics/claude-code/issues/18454 https://github.com/anthropics/claude-code/issues/19252 https://github.com/anthropics/claude-code/issues/19477 https://github.com/anthropics/claude-code/issues/19500 https://github.com/anthropics/claude-code/issues/19739 https://github.com/anthropics/claude-code/issues/20154 https://github.com/anthropics/claude-code/issues/22503 https://github.com/anthropics/claude-code/issues/23032 https://github.com/anthropics/claude-code/issues/25166 https://github.com/anthropics/claude-code/issues/26533 https://github.com/anthropics/claude-code/issues/28006 https://github.com/anthropics/claude-code/issues/32048 https://github.com/anthropics/claude-code/issues/32163 https://github.com/anthropics/claude-code/issues/32198 https://github.com/anthropics/claude-code/issues/32523 https://github.com/anthropics/claude-code/issues/32554 https://github.com/anthropics/claude-code/issues/32659 https://github.com/anthropics/claude-code/issues/34197 https://github.com/anthropics/claude-code/issues/34358 https://github.com/anthropics/claude-code/issues/38072 https://github.com/anthropics/claude-code/issues/38193 https://github.com/anthropics/claude-code/issues/40289 https://github.com/anthropics/claude-code/issues/42046 https://github.com/anthropics/claude-code/issues/42295 https://github.com/anthropics/claude-code/issues/44803 https://github.com/anthropics/claude-code/issues/44960 https://github.com/anthropics/claude-code/issues/47598 https://github.com/anthropics/claude-code/issues/47101 https://github.com/anthropics/claude-code/issues/48293 https://github.com/anthropics/claude-code/issues/50027 https://github.com/anthropics/claude-code/issues/50734 https://github.com/anthropics/claude-code/issues/51174 https://github.com/anthropics/claude-code/issues/53223 https://github.com/anthropics/claude-code/issues/54033 https://github.com/anthropics/claude-code/issues/57392 https://github.com/anthropics/claude-code/issues/60339 https://github.com/anthropics/claude-code/issues/70125 https://github.com/anthropics/claude-code/issues/70420 https://github.com/anthropics/claude-code/issues/71577 https://github.com/anthropics/claude-code/issues/72651 https://github.com/anthropics/claude-code/issues/74414 https://github.com/anthropics/claude-code/issues/77544

Proposed Solution

I don't know if a solution is possible.

Maybe we need custom system prompts (https://github.com/anthropics/claude-code/issues/7571), or wait for another generation or architecture of AI models. Current models fail to obey instruction hierarchies (e.g. system vs user prompt separation) https://github.com/yilin-geng/llm-instruction-conflicts, direct instructions https://github.com/THU-KEG/AgentIF, and even reasoning models fail to follow instructions https://github.com/ykwon0407/reasonIF.

In any case, I wold like to see periodic, official statements what is the measurable state of instruction following in Claude Code.

Alternative Solutions

Periodically remind Claude Code during a conversation to re-read CLAUDE.md https://github.com/anthropics/claude-code/issues/2901#issuecomment-3065931466.

Even this is sometimes ineffective, and I need to explicitly instruct Claude Code to behave in a certain way, despite this being already described in CLAUDE.md.

Priority

High - Significant impact on productivity

Feature Category

Performance and speed

Use Case Example

Claude Code (all generation models, including top models like Opus 4.5, 4.6, 4.7, Fable 5) tend to ignore CLAUDE.md instructions. Sometimes this happens at the beginning of a conversation (see a reproduction repository https://github.com/marcindulak/claude-fails-to-follow-claude-md), but more often during longer conversations.

The problem is probably not in the particular wording of CLAUDE.md, since: 1) sometimes the instructions are followed, 2) there are no rules how to deterministically improve CLAUDE.md so it is followed

The failure of Claude Code to follow CLAUDE.md instructions requires a human to be constantly present in the loop and defies the premise of autonomous agents.

Additional Context

See a similar issue for https://github.com/google-gemini/gemini-cli/issues/13852

_No response_