Clarke Teach

Mode: Create a Policy

Triggers: "always", "never", "from now on", "the rule is", "add a policy"

1. Ask for:

   • Policy content: The rule in clear, actionable language
   • Owner: Who owns this (default: current user)

2. Call clarke_create_policy

3. Explain: "Policy created in draft status. It may need approval before influencing agent behavior. Use /clarke-review to check."

Mode: Submit a Correction

Triggers: "that was wrong", "actually it should be", "incorrect"

1. Ask for:

   • What was wrong: The specific bad answer
   • What it should be: The correct answer
   • Request ID (if available): From the trace_id of the bad query

2. Call clarke_feedback with accepted: false and detailed notes

3. Explain: "Correction submitted. If enough similar corrections accumulate, CLARKE will propose a behavioral directive. Check /clarke-review periodically."

Mode: Positive Feedback

Triggers: "that was great", "perfect", "exactly right"

1. Identify the request being praised
2. Call clarke_feedback with accepted: true
3. Confirm: "Positive feedback recorded. This helps CLARKE learn which patterns work."

Tools

Example

User says: "we decided to use structlog instead of stdlib logging because it gives us structured JSON output"

Agent does:

1. Recognizes as a Decision
2. Extracts title, rationale
3. Asks: "Were there alternatives considered?"
4. User: "considered stdlib logging and loguru"
5. Calls clarke_create_decision
6. Confirms: "Decision recorded: 'Use structlog for logging'"