---
name: "Skill runtime — load, execute, unload + approval field"
description: "Use when a skill runtime must load, execute, unload, and name what requires approval."
---

# Skill runtime — load, execute, unload + approval field

## When
Use when writing or reviewing a skill's six fields — especially the approval field people skip.

## Stop
A skill without approval boundaries is a bot that makes consequential decisions alone. One skill at a time.

## Steps
1. Scan all skills (bundled, project, global, plugins, community) as YAML metadata so the system prompt stays lean
2. Query arrives; tools already known
3. Model picks the skill from **descriptions** (not the user)
4. Activate: metadata → SKILL.md → scripts only if needed
5. Inject instructions, scope permissions
6. Execute with tools
7. Deliver output, unload, next skill
8. When to use it

## Load
Skill YAML metadata scan. SKILL.md body on activate. Approval field required.