Improve from evidence

Use eval evidence to improve a skill. Workbench edits package source, proves the candidate with eval evidence, and records lineage for review.

See Evaluation basics for eval setup, Results for evidence inspection, Review candidate changes for candidate review, and Rerun proof evals for another proof attempt.

When to improve

Improve after below-perfect, failed, or reviewed graded evidence:

workbench eval resultsworkbench eval listworkbench eval show RUN_IDworkbench skill improve --dry-runworkbench skill improveworkbench eval results

Perfect-only projects need better cases, stricter grading criteria, or higher samples before improvement is useful. Add or tighten eval cases instead of asking for cosmetic source changes. After changing eval source, use workbench eval list or workbench eval results --eval eval-vN when you need to compare old and current score meanings.

Candidate and proof

workbench skill improve needs one package version and one agent. Use --versions and --agents when defaults would select more than one entry.

workbench skill improve --versions current --agents defaultworkbench skill improve --budget 2workbench skill improve --cloud

The improver edits package source outside .workbench/**, runs proof evidence, and switches only when the proof succeeds and beats the current version. Improvement changes the skill package, not the grading standard.

Next steps

• Review candidate changes explains skill versions, skill diff, and evidence inspection for a candidate.
• Rerun proof evals explains --rerun, retry, and higher-sample proof runs.
• Publish explains source publication after a candidate is ready.