💬
← All posts

A Ten-Minute Audit for Any Skill

2026-10-23

You do not need a rubric to find the common defects. You need ten minutes and five questions.

This checklist is derived from the corpus patterns in this batch. It is ordered by impact per minute.

Minute 0–2: Routing

Read only the description (not the body) and answer:

  • Does it name the concrete situations that should trigger it?
  • Does it name at least one adjacent request that should not?
  • Does it say what the user gets?

If the exclusion is missing, that is the highest-leverage edit you can make. Remember: 25.6% of skills have a boundary in the body, and the router never reads the body.

Minute 2–5: Workflow

Read the workflow and count the steps.

  • Zero numbered steps: is the sequence obvious? If not, add it.
  • More than fifteen: are these really one workflow, or several skills?
  • Does any step ask for missing input instead of guessing?
  • Does any step verify the result?

The median skill has six steps. That is a reasonable altitude: enough to remove ambiguity, not so much that the agent is executing a script.

Minute 5–7: Output and failure

  • Is there an output section? Does it define the artifact, its structure, and what to report back?
  • What happens if an API call fails, a file is missing, or a test does not pass?

Only 7.8% of the corpus describes failure handling. One sentence is enough: "If the response is not JSON, stop and report the status code."

Minute 7–9: Security

Three questions, in order:

  1. What can it execute? If Bash appears, what commands?
  2. What can it delete or overwrite? Is the path scoped? Is there a confirmation?
  3. What can it leak? Does it read files, environment variables, or credentials it does not need?

    4. Only 1.8% of skills require confirmation before irreversible actions. If yours can force-push, delete, migrate, or publish, add the gate.

    Minute 9–10: Structure

    • Is the body longer than necessary? Move detail into references/ if it is only needed sometimes.
    • Are there assets or scripts the skill references but does not include?
    • Does the file contain a .DS_Store, a build artifact, or a placeholder URL in executable code?

    What the audit will not catch

    Domain correctness, trigger quality in practice, and whether the skill solves a real problem. Those require using it. The audit catches the defects that are visible before anyone runs it — which, across this corpus, are the majority.

    The one-line version

    If you remember nothing else: put the trigger and the exclusion in the description, put a verification step in the workflow, and put a confirmation gate in front of anything irreversible.