Synthesize

Use this skill to turn source material into official skill eval manifests and explicit APO datasets.

In this repository, authored eval cases live in evals/evals.json, optional eval assets live in evals/files/, and APO training data lives in explicit datasets/train.jsonl and datasets/val.jsonl files. Work compute-first. If labels, expected outputs, assertions, or JSONL rows depend on derived values, plan the computation before writing them. Synthesize only after the ground truth is provided or derived.

When to use this skill

• The workflow needs evals/evals.json for a markdown prompt or skill.
• The workflow also needs explicit train.jsonl and val.jsonl files before handing off to a downstream optimizer.
• The user has examples, CSV or JSON data, tables, schemas, business rules, or source notes from an approved research brief.
• The agent should convert known source material into high-quality eval cases.
• The agent should convert grounded source rows into explicit APO datasets without blurring them together with authored eval manifests.
• The expected outputs are not directly written down and must be derived from raw fields.
• The agent needs explicit transforms such as map, filter, reduce, grouping, joins, sorting, normalization, or rule evaluation before authoring eval rows.
• The agent should add synthetic rows only after grounded or computed rows establish the target task pattern.

If no credible source material exists yet, use a primary-source research workflow first instead of guessing.