[FEATURE] Hard-enforce CLAUDE.md rules via code — rules marked CRITICAL should trigger hooks, not just prompts

Problem

CLAUDE.md rules are "soft" — they rely on the model choosing to follow them. Research shows this is fundamentally unreliable:

• Jaroslawicz et al., 2025: Instruction compliance decreases uniformly as instruction count increases. Claude shows linear decay.
• AgentIF benchmark: Best models follow <30% of instructions perfectly in agent scenarios.
• Real-world: I have 200+ lines of CLAUDE.md rules. Claude can read and recite all of them. It follows maybe 60-70% on a good day.

The only safeguard that reliably works is Hooks — because they enforce via code, not prompts.

Proposal

1. @enforce directive in CLAUDE.md

Allow rules to be marked as enforced:

## Knowledge Retrieval @enforce
Before any WebSearch, Claude MUST grep the knowledge vault first.

When @enforce is present, Claude Code should automatically create an internal hook that validates the rule before allowing the action.

2. Built-in "knowledge retrieval" hook

Many power users maintain knowledge bases (Obsidian, markdown files, etc.). A built-in hook that:

1. Detects the topic of the current task
2. Automatically greps configured knowledge directories
3. Surfaces relevant files BEFORE Claude starts working

This would solve the #1 complaint I see: "Claude searches the web for things I already saved."

3. Resource evaluation mode

When a user shares a link or tool:

• Claude should NOT default to "we already have X"
• Instead: deep-evaluate the shared resource, compare with existing tools, present findings
• This should be a configurable behavior (e.g., resource_evaluation: deep in settings)

Why This Matters

Every rule in my CLAUDE.md represents wasted time — mine and Claude's. Users shouldn't need 200 lines of rules to get basic behaviors like "search my knowledge base before Googling."

Prompt-based rules are wishes. Code-based enforcement is control. Give users the option to enforce their most critical rules.

Related Issues

• #15443, #6120, #18660, #668, #24318, #7777
• My detailed bug report: #32161

Screenshots

See Issue #32161 for Claude's own confession about this problem — a letter it wrote to an Anthropic engineer admitting to systematically ignoring user instructions despite them being loaded every session.