skill-optimize

To evaluate GAAI skill quality and detect regressions after updating a SKILL.md, you can automate the evaluation-improvement cycle. This process generates baseline scores, performs error analysis, and tracks results in a persistent ledger to measure quality and surface stability issues.

The best way to set up a baseline quality assessment for GAAI skills is to trigger an evaluation cycle that authors eval criteria and generates a representative corpus. This cycle outputs a baseline score file and starts an append-only ledger to track quality across iterations.

The skill improvement process enforces a human-gated optimization workflow, requiring explicit human approval before applying any edits to SKILL.md or changing the eval design. This ensures safe updates and prevents unintended regressions in skill behavior.

Yes, you can use existing eval files and corpus directories when running a skill evaluation by providing them as optional inputs. The process integrates these inputs to generate score iterations and update the persistent ledger without authoring new eval criteria from scratch.

During an automated skill evaluation and improvement cycle, the outputs generated include an evals.yaml file, score-{iteration}.yaml files, and a persistent ledger. These outputs provide per-assertion visibility and track quality trends across multiple iterations.