compass-optimize-instructionlisted

# compass-optimize-instruction Generate 5 framings, score on fixed test cases, select, critique the winner. The guard enforces the candidate count, the framing set, the tie-break, and the checklist size in code. `GUARD="python3 ${CLAUDE_PLUGIN_ROOT}/scripts/compass.py"` ## Preferred path: the workflow When the Workflow tool is available, run `${CLAUDE_PLUGIN_ROOT}/workflows/optimize-instruction.js` (args `{ taskDescription, testCases:[{input, expectedOutcome}] }`). It dispatches one `instruction-candidate` agent per framing in parallel, scores each against the exact test cases, selects with the framing tie-break, and critiques the winner. ## Inputs - a recurring task description - **3-5 representative real test cases** `{input, expectedOutcome}`. **Source them from real prior inputs first; only construct if none exist. MUST NOT fabricate test cases or adjust expected outcomes.** ## Process 1. **Generate exactly 5 candidates — one per APE framing:** rule-based, example-based, definition-based, question-based, chain-of-thought-based. Each commits fully to its framing (no blending). 2. **Score each candidate against the exact provided test cases** (pass count). 3. **Select the highest score.** Validate the count, framings, and tie-break: ``` echo '{"candidates":[ {"framing":"rule-based","score":4}, {"framing":"example-based","score":4}, {"framing":"definition-based","score":3}, {"framing":"question-based","score":2}, {"framing":"chain-of-thought-based"