Skip to content

Latest commit

 

History

15 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 

Repository files navigation

skill-composition

A skill that teaches an agent how to write, review, and score other agent skills.

What This Repository Holds

The skill lives in skills/skill-composition/. It contains:

  • SKILL.md — the six-step process to write a skill.
  • references/rubrics.md — sixteen rubrics (R1-R16) that score a skill, with pass lines and an N/A rule.
  • references/file-structure.md — the folder layout, the file limits, and the naming rules.
  • references/reference-files.md — the rules for reference files.
  • references/placement.md — where skills live in a project: owned skills in skills/, vendored skills in the untracked load path.
  • references/skill-families.md — the rules for sibling skills: the router pattern, disjoint triggers, and one wording for shared steps.
  • references/project-audit.md — the five-step audit workflow for every skill in a project.
  • README.md — the usage notes for the skill itself.

What the Skill Does

An agent with this skill can:

  • Write a new skill with one job, a searchable description, and specific instructions.
  • Review an existing skill and report each problem with a rubric ID, the text with the problem, and a corrected version.
  • Score a skill against the rubrics and give a pass or fail result.
  • Split a skill that mixes more than one job.
  • Organize the skills of a project: owned skills in skills/, vendored skills in the untracked load path, and a policy in AGENTS.md.
  • Audit every skill in a project and deliver one report with scores, problems, and a correction plan.

Why You Use It

Skills fail for known causes. A skill with two jobs does neither job well. A vague description never triggers. A vague instruction gives different output on each run. The rubrics in this skill catch each of these causes before you ship the skill.

The skill also sets language rules for the skills it writes: short sentences, imperative steps, one word for one meaning, and no optional modals. Models read "should" as optional. These rules write "must". This makes the instructions hard to misread, for models and for people.

The structural rules come from the Skills Directory documents "Writing Effective Skills", "Skill File Structure", and "Using Reference Files".

How to Install It

Install with the skills CLI:

npx skills add markab21/skill-composition

Or copy skills/skill-composition/ into the skills folder of your agent.

License

MIT. See the LICENSE file.

About

A skill that teaches an agent how to write, review, and score other agent skills.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors