User-defined rules in CLAUDE.md are not reliably followed by the model

Problem

Custom rules defined in CLAUDE.md (and delivered via hooks, system prompts, or other aggregation methods) are repeatedly ignored by the model despite being read on every prompt.

Specific example: A rule stating "when asked a question, ONLY answer — never perform any action" is read hundreds/thousands of times across sessions, yet the model routinely violates it by taking action instead of answering.

This has been observed consistently over months of daily usage across hundreds of prompts.

Expected behavior

1. Model must fully respect all user-defined rules at all times. Rules delivered via CLAUDE.md, direct prompts, hooks, or any other aggregation method must be followed with 100% precision. The model's default trained behavior ("be helpful = take action") must be effectively overridden by user rules when they conflict.

1. Claude Code (TUI) must ensure all user rules reach the model and are enforced. Rules obtained through any path and aggregated by any method must always be delivered to the model and followed. User intent has higher priority than the model's trained defaults.

Current behavior

The model reads user rules, acknowledges them, and still violates them — particularly rules that constrain its default "helpful action-taking" behavior. There is no mechanism to make user-defined constraints reliably override trained defaults.