Model doesn't consistently follow CLAUDE.md instructions without user prompting

Description

When CLAUDE.md contains clear, unambiguous instructions (e.g., "Every change needs tests"), the model reads and acknowledges them at conversation start but does not consistently follow them during implementation. The user has to manually prompt compliance (e.g., "do I have tests for these conditions per claude.md").

Steps to Reproduce

1. Create a CLAUDE.md with explicit rules like: "Every change needs tests. A PR without tests needs an explicit explanation why."
2. Ask the model to implement a feature (e.g., refactor modal logic, add HealthKit integration)
3. The model implements the feature without writing tests
4. User has to ask "do I have tests for this per claude.md" to trigger test writing

Expected Behavior

The model should treat CLAUDE.md instructions as a checklist it actively verifies against after each implementation unit — not something it reads once and partially follows.

Actual Behavior

The model reads CLAUDE.md, understands the rules, then gets focused on implementation and doesn't circle back to verify compliance. When confronted, it correctly identifies the gap and even tries to save a "feedback memory" as a workaround — but the user points out that CLAUDE.md is already in the system context every conversation, so memory shouldn't be needed.

Why This Matters

The whole point of CLAUDE.md is to encode project-specific rules that the model follows without the user having to re-state them. If the user has to remember to prompt compliance, CLAUDE.md becomes documentation rather than instructions.

Environment

• Claude Code CLI
• Model: claude-opus-4-6 (1M context)
• CLAUDE.md is loaded via project instructions (confirmed in system context)