skill-improve

To iteratively improve a skill, you run a static test baseline, diagnose failing and warning checks, apply targeted before/after edits, and retest to decide whether to keep or revert changes based on measurable score improvements.

Diagnosing failing rubric compliance checks involves running static tests to pinpoint specific gaps like missing frontmatter fields, insufficient phase structure, missing verdict keywords, or absent ask-before-write language in your SKILL.md file.

Yes, you can automatically rewrite skill body instructions to fix static analysis warnings by applying targeted before/after edits to failing sections, but only after user approval and using an allowed-tool-constrained write step.

A test-driven improvement loop for skill maintenance is a process where you execute static and category checks, propose targeted edits to rewrite only what fails or warns, and then retest to ensure changes yield measurable improvements.

To safely revert skill changes when retesting shows lower scores, the improvement workflow keeps improvements only if the combined score improves, otherwise it offers guidance to revert the applied edits.

Yes, you need user approval before writing changes during skill improvement, as the workflow requires an allowed-tool-constrained write step that executes only after explicit approval to modify the SKILL.md file.