Skip to main content
The fastest way to author a skill is to use the bundled skill-creator skill from Claude Code inside a sandbox terminal. It walks you through the folder layout, writes a well-formed SKILL.md, scaffolds scripts, and tests it end-to-end.

​
Folder layout

my-skill/
├── SKILL.md          # Required — what the AI reads
├── scripts/          # Optional — runnable helpers
│   └── process.py
└── references/       # Optional — longer-form docs
    └── guide.md

​
SKILL.md

Frontmatter + prose instructions: 
---
name: my-skill
description: Brief one-line description — this is what the AI sees in <available_skills>
---

# My Skill

## When to use
Describe the task signal that should make the AI reach for this skill.

## Workflow
1. ...
2. ...

## Scripts
- `scripts/process.py` — what it does, how to call it.

## Output format
Be explicit about where files should land and what shape the user expects.
Keep SKILL.md short. The AI reads it once per invocation — walls of text eat context. Use references/ for long docs, and tell the AI in SKILL.md when to consult them.

​
Scripts

Scripts live under scripts/ and are invoked via bash_tool. Typical shape: 
# scripts/process.py
import argparse, sys, json

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--input", required=True)
    ap.add_argument("--output", required=True)
    args = ap.parse_args()
    # ...
    print(json.dumps({"ok": True, "path": args.output}))

if __name__ == "__main__":
    main()
Design for unix: one job, parseable output, non-zero exit on failure.

​
Testing

  1. Drop your skill into /mnt/skills/public/my-skill/ inside a sandbox (bind-mount or temporary copy).
  2. Start a fresh chat.
  3. Ask for the thing — the AI should pick the skill automatically.
  4. Iterate on SKILL.md if the AI keeps reaching for the wrong tool.

​
Publishing

​
Managed Yambr

Custom skills are stored in the Settings Wrapper registry. Zip your skill, upload it via the dashboard (or let the AI do it via the settings-manager skill), and it becomes available on your account.

​
Self-hosted

  • Bake into the image. Drop the folder under skills/public/my-skill/ in your fork of the repo and rebuild. Everyone on that deployment gets it.
  • Bind-mount at runtime. Use the Settings Wrapper’s ZIP-upload flow, or mount a host directory into /mnt/skills/user/my-skill/.