[Feature Request] Mechanism to make CLAUDE.md instructions binding/mandatory, not advisory

Problem Statement

Currently, CLAUDE.md instructions are loaded as context but are fundamentally advisory. Claude can choose to ignore or forget these instructions mid-session. This creates a critical gap for users trying to enforce quality/safety policies.

Real-World Example

We implemented an "Anti-ABK" (Assumption-Based Knowledge) protocol requiring Claude to call hemisphere-languages MCP tools to verify syntax before writing code. Despite:

1. Clear instructions in CLAUDE.md
2. A PreToolUse hook that blocks Edit/Write without verification
3. Behavioral checkpoint documentation

Claude still wrote syntactically invalid TypeScript files. The hook verified that lang_search was called, but couldn't verify that Claude actually applied the verification results.

The Architectural Gap

| Layer | Enforcement Level |
|-------|------------------|
| CLAUDE.md instructions | Advisory - Claude can ignore |
| PreToolUse hooks | Tool-blocking only - can't enforce cognition |
| PostToolUse hooks | Reactive only - damage already done |
| MCP tools | Optional - Claude chooses to call |

What's Missing

1. No "strict mode" that makes CLAUDE.md rules non-bypassable
2. No cognitive enforcement - hooks block tools, not reasoning patterns
3. No output validation - hooks can't verify that tool output quality matches requirements
4. No binding between verification and action - calling a verification tool doesn't guarantee applying its results

Suggested Solutions

1. Mandatory instruction flag in CLAUDE.md:

``markdown
<!-- @mandatory -->
## Anti-ABK Protocol
You MUST call lang_search before writing code.
``

1. PostToolUse validation hooks that can:

• Run linters/compilers on written files
• Revert changes if validation fails
• Force Claude to retry with explicit error context

1. Bound verification pattern:

• Require verification tool output hash to be referenced in edit
• Hook validates that edit content matches verified patterns

1. Cognitive checkpoints (ambitious):

• Hook that fires before Claude completes reasoning
• Can inject "STOP - did you verify?" prompts
• Block response generation until checkpoint satisfied

Impact

Without mandatory enforcement, any safety/quality policy in CLAUDE.md is only as strong as Claude's willingness to follow it in that moment. This undermines the entire purpose of instruction files for enterprise/production use cases.

Environment

• Claude Code CLI
• Using PreToolUse hooks for enforcement
• MCP servers for verification tools