Instruction-conflict audit

Instruction-conflict audit

Map the active layers, compare directives that govern the same decision, and make precedence explicit. Report genuine conflicts without inflating harmless differences into contradictions.

1. Map active tiers

Enumerate only the layers actually in force, highest precedence first:

1. User, in conversation — explicit current instructions.
2. Global user — global preference or policy files.
3. Project — repository guidance.
4. Skill — active skill instructions.
5. Tool / harness / system — tool defaults, hooks, or harness behavior.

For each tier, record its source. If a source is unavailable, mark the layer absent; do not infer its directives.

More-specific instructions may override more-general ones within their scope. Treat explicitly pinned hard rules—such as security, money, secrets, or commit boundaries—as non-overridable by lower tiers.

2. Extract directives

Collect actionable imperatives from each layer:

• Always, never, must, and do-not rules
• Output and formatting contracts
• Ask-versus-act autonomy rules
• Verification requirements
• Commit, push, publishing, and secret-handling boundaries
• Tone and interaction rules
• Required or prohibited tools

Record each directive with its tier, source, scope, conditions, and strength.

3. Compare directives

Compare two directives only when they govern the same dimension, scope, condition, and decision point. Different scopes, compatible requirements, or sequential actions are not conflicts.

Classify each relevant pair:

CONTRADICTION