firehose> #llmops

Skill Authoring Checklist

A shared rubric for judging and improving an individual agent skill, offered to fill what the source calls the missing piece of the skills ecosystem: there is "no shared framework for looking at a skill and making it better," so people can't tell a good skill from a bad one and don't get the results skills promise ("skill hell"). The checklist has four pillars, applied in order both to write a skill and to audit an existing one:

  1. Trigger — how the skill is invoked, and the user-vs-model decision it forces (Skill Invocation Trigger).
  2. Structure — the internal layout: steps + reference, and keeping SKILL.md minimal (Minimal Skill Surface).
  3. Steering — getting the agent to actually do what you want, chiefly via Leading Words and, where needed, Step Isolation for leg work.
  4. Pruning — a final smallness pass hunting DRY violations, sediment, and no-ops (Skill Pruning).

The rubric is domain-agnostic authoring craft, distinct from the vault's composition/reuse cluster: Skill-Driven Loop Development is about pre-validating skills before looping over them and Reusable Workflow Library/Agentic Distribution about cataloging and sharing them — this checklist is one altitude lower, about the internals of a single skill.

Claims


Linked from