What problem does it solve? Manually rewriting prompt templates is slow and unmeasurable. This Skill automates prompt improvement by running Agent Lightning APO (Automatic Prompt Optimization) against explicit train and validation datasets, returning a scored, optimized markdown prompt with a JSON report. ## Core Features & Use Cases - Automated Prompt Optimization: Runs APO or VERL beam-search optimization over a markdown prompt template using JSONL train/val datasets, with deterministic, custom, or LLM-judge scoring modes. - Safe, Non-Destructive Output: Returns the optimized prompt in JSON/stdout by default; file writes happen only when explicitly requested via output, report, or in-place flags. - Offline Fallback: When no inference model is configured or rate limits hit, it returns a manual_followup payload with a ready-to-use model prompt so an agent can complete the optimization without an API token. - Use Case: You have a support-intent classifier prompt and a CSV of labeled examples. Generate train.jsonl and val.jsonl, then run the optimizer to get a measurably better prompt plus a dashboard URL for inspecting rollouts. ## Quick Start Ask the agent to optimize your markdown prompt with /trainer-optimize by providing the prompt file path along with explicit train and validation JSONL dataset paths.
