prompt-optimizer-looplisted

# Prompt Optimizer Loop Improve a prompt the way engineers improve code: one change at a time, measured against a locked test, keep winners, revert losers. Interactive — the user approves each round. ## Step 0 — Collect the three inputs 1. TARGET: the prompt to improve (verbatim). 2. TEST INPUTS: 2-3 realistic inputs that would hit this prompt in production. If the user has none, help draft them, then confirm. 3. CHECKLIST: 3-6 binary yes/no quality checks (e.g., "output is valid JSON", "response under 100 words", "never invents a source"). If missing, draft candidates from the user's complaints and get approval. Once approved, the checklist is LOCKED — it cannot change during the loop. Changing it starts a new loop. ## Step 1 — Baseline Run the TARGET prompt against every test input. Score each output against every checklist item. Record the baseline as passes/total (e.g., 9/15 across 3 inputs × 5 checks). ## Step 2 — One mutation Identify the checklist item failing most often. Propose exactly ONE change to the prompt targeting that failure (add a constraint, reorder, add an example, tighten wording). State the hypothesis: "Adding X should fix check Y because Z." Never bundle multiple changes — attribution dies. ## Step 3 — Re-test Run the mutated prompt against ALL test inputs (not just the failing one). Score against the full locked checklist. ## Step 4 — Keep or revert - Score improved AND no previously-passing check broke → KEEP. New version becomes current. -