| name | skill-improvement |
| description | Add, refactor, validate, or retire an agent skill. Use when implementation reveals a durable rule, when a skill is too large or stale, or when the skill catalog changes. |
Skill improvement
When to use
- A change reveals a durable, non-obvious repository rule.
- A skill contains duplicated, dated, or contradictory instructions.
- A new repeatable workflow needs an agent entry point.
- A workflow or source path invalidates a skill.
When not to use
Do not use this for session notes, issue status, completed-work lists, or a
one-off command that has no future value.
Procedure
-
Identify the one workflow or invariant the skill should own.
-
Search docs/skills/INDEX.md and existing skills for an existing owner.
-
Update the existing skill instead of creating a duplicate.
-
If new, create docs/skills/<lowercase-kebab-case>/SKILL.md.
-
Add only this front matter:
---
name: <directory-name>
description: >-
<capability>. Use when <specific activation conditions>.
---
-
Start with When to use, When not to use, and an actionable Procedure.
-
Add Red flags and Verification when they improve safe execution.
-
Keep the entry file under 500 lines; move large references to an on-demand file.
-
Link to canonical facts instead of copying them from another document.
-
Add or update the entry in docs/skills/INDEX.md.
-
Remove stale text that contradicts the new rule.
-
Run the documentation checks before handoff.
Content rules
Include specific activation language, imperative steps, required source files or
commands, repeatable failure modes, and evidence required before completion.
Exclude dated incident narratives, issue or pull-request numbers, session logs,
status percentages, duplicated repository-wide policy, and advice with no
observable action.
Canonical ownership
Common rationalizations
- “This fact is obvious.” If it caused a failure or workaround, document it once.
- “I will update the skill later.” Capture durable learning in the same change.
- “A second skill is clearer.” Search for an existing canonical owner first.
Red flags
- The skill has no specific activation conditions.
- The entry file contains a long reference catalog instead of a procedure.
- The same rule appears in multiple skills.
- The skill describes a deleted source path or unverified command.
Verification
python3 scripts/check-skill-docs.py
pre-commit run --all-files
Also verify that the skill is indexed, repository-relative links resolve, and
the entry file stays within the size budget.
