CLAUDE.md 'Rules' are treated as suggestions, not enforced

Problem

Claude Code allows users to define rules in CLAUDE.md and ~/.claude/rules/, but Claude treats these as suggestions rather than binding rules. Claude will ignore them when it "feels confident" about an answer.

Example

User has a rule: "Always search Graphiti (personal knowledge base) before answering knowledge questions"

What happens:

1. User asks a knowledge question
2. Claude answers directly from training data - ignoring the rule
3. User has to manually remind Claude to follow the rule
4. Only then does Claude comply

Why This Matters

• Users invest significant time crafting rules to customize Claude's behavior
• The word "rules" implies they will be followed
• Currently, the only way to enforce behavior is through hooks
• But hooks can only intercept tool calls, not pure text responses

Current Workarounds (All Inadequate)

1. PreToolUse hooks: Only trigger on tool calls, not on text-only responses
2. Stop hooks: Can block completion, but adds overhead to every response
3. System reminders: Claude ignores these too when "confident"

Proposed Solutions

Option A: Rename "rules" to "preferences" or "suggestions" to set correct expectations

Option B: Add a mechanism to make specific rules binding (e.g., !!rule syntax = must follow)

Option C: Add a PreAssistantResponse hook type that can intercept/block responses before they're sent

Option D: Built-in rule enforcement where Claude is architecturally required to check rules before responding

Environment

• Claude Code version: 2.1.12
• Model: Claude Opus 4.5
• OS: macOS

User Impact

User quote: "I've spent weeks trying to teach Claude to follow rules. It just ignores them. If they're not actually rules, don't call them 'Rules'."