What problem does it solve? Turning raw source material like CSV rows, tables, schemas, and business rules into high-quality evaluation cases and training datasets is error-prone, especially when expected outputs must be computed rather than copied. This Skill structures that conversion so eval manifests and APO datasets stay separate, verified, and leakage-free. ## Core Features & Use Cases - Eval manifest authoring: Produces official evals/evals.json cases with realistic prompts, expected outputs, optional files, and objective assertions. - APO dataset generation: Writes explicit datasets/train.jsonl and datasets/val.jsonl files with evaluator-facing fields kept out of the prompt-visible input path. - Verifier-backed synthesis: Computes derived outputs with explicit transforms (map, filter, reduce, joins, rule evaluation) and independently verifies synthetic rows before keeping them. - Use Case: Given grounded support-ticket examples and a prompt file, generate the eval manifest plus train/validation splits with a defensible holdout strategy for a downstream optimizer. ## Quick Start Use the trainer-synthesize skill to build evals/evals.json plus train.jsonl and val.jsonl datasets from my grounded CSV rows for the prompt at prompts/classify_support.md.
