What problem does it solve? Optimizing prompt files by hand is error-prone: placeholders get renamed, evaluator fields leak into prompt text, and scoring modes are guessed inconsistently. This Skill provides a structured orchestration contract that runs the full trainer loop (research, dataset synthesis, optimization, election, validation, write-back) against prompt-like targets with enforced gates and resumable workspace state. ## Core Features & Use Cases - End-to-end prompt training loop: Initializes a .trainer-workspace/ next to the target, enforces an engineering review checkpoint, runs optimization passes, handles manual follow-up branches, and writes back only validated winning candidates. - Prompt-specific guardrails: Infers judge mode (llm_judge vs deterministic) from dataset row shape via a routing table, preserves template placeholders exactly, and keeps evaluator-only fields (expected, reference, criteria, scoring) out of the prompt-visible text. - Blocker-first resumability: Stops with a named blocker and a resumable workflow-status.json checkpoint when reviews, datasets, or tracked artifacts are missing or inconsistent. - Use Case: Given prompts/summarize.prompt.md with no training assets, the loop derives the workspace, requires the review checkpoint, synthesizes train/validation datasets, optimizes the prompt, and writes back the winner only after validation passes and placeholders are confirmed unchanged. ## Quick Start Run the trainer loop for prompts/summarize.prompt.md, initializing the workspace and enforcing the engineering review checkpoint before any optimization pass.
