---
name: refine-skill
description: 'Audit and slim a SKILL.md: invocation mode, description, body, references.'
disable-model-invocation: true
---

# Refine skill

Apply [agent-copy](../agent-copy/SKILL.md) in full; Read it first if not already loaded. Its pointer rules audit the description, and its skill-mechanics section audits the invocation mode. Checks specific to skills:

- **Check the invocation mode first.** A wrong mode invalidates how the description should read.
- Model-invoked descriptions follow the template `Use when the user asks to …`, in the user's words.
- Examples are concrete and adaptable, not generic.
- Work re-derived across runs (a converter, a query, a validation) becomes `scripts/`.
- Verify tool names, flags, and paths still exist: a skill caches the world and the world moves.

When quality is contested or the description mis-triggers in practice, go empirical: run representative prompts with and without the skill and compare.
