Skill management
What it does
Creates, inspects, and deletes custom managed skills. The meta-skill that lets you extend the assistant. Companion to Your first skill.
Setup
None. Works immediately.
Permissions
- Create and delete are high-risk and always require approval. That is the privilege-escalation guard.
Common prompts
| You say | What happens |
|---|---|
| “Build me a skill that checks Hacker News for trending posts” | Scaffolds a managed skill |
| “Create a skill for the project's deploy checklist” | Domain-specific skill |
| “Delete the Hacker News skill” | Removes a custom skill |
| “Show me my custom skills” | Lists managed skills |
Configuration
- A skill folder: SKILL.md plus optional references/*.md and scripts/* (python3 or bun).
- Companion scripts are files the assistant runs from the terminal. Skills built this way do not register executable TOOLS.json (that ships with built-in and community-installed skills).
- On a Linux hatch, managed skills live under that instance's workspace
skills/directory, not a global ~/.cadence for every assistant.
Tips
- The assistant can draft, test in a sandbox, and persist. You still approve the write.
- Review generated code before Allow.
- When a scripts/ file is stored, it is code that already ran successfully, so reuse does not rewrite from scratch.
- Live inventory:
assistant skills list. See Self-improving skills.
Related: Features, Skills reference.