# New Skill

> Scaffold a new skill in this skills repository — copies the template, enforces Agent Skills spec conventions, refreshes the catalog and marketplace manifest, and relinks agent skill directories. Use when asked to create, add, or scaffold a skill (새 스킬, 스킬 만들어줘, add a skill).

- Skill: `soonjune/new-skill` (Agent Skill)
- Install (CLI): `npx skillmds@latest add soonjune/new-skill`
- Raw SKILL.md: https://api.skillmd.com/api/skills/soonjune/new-skill/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT
- Author: soonjune (https://skillmd.com/u/soonjune)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/soonjune/new-skill

---


# New Skill

Create a skill following this repository's conventions. Skills here serve multiple agents (Claude Code, OpenClaw, Hermes) through the Agent Skills open standard, wired up by symlinks.

## Steps

1. **Name it.** Lowercase letters, digits, hyphens; max 64 chars; no leading/trailing/consecutive hyphens. The directory name must equal the frontmatter `name`.
2. **Scaffold.** From the repo root: `cp -r template skills/<name>`, then edit `skills/<name>/SKILL.md`.
3. **Write the description** (max 1024 chars, one line, third person): what it does AND when to trigger it, with the concrete phrases a request would contain, in every language the owner actually uses.
4. **Write the body.** Imperative instructions, under 500 lines. Long reference material goes to `references/*.md` with explicit "load this when..." pointers; runnable helpers go to `scripts/`.
5. **Cross-agent rules.** Shared skills stick to standard frontmatter fields (`name`, `description`, `license`, `compatibility`, `metadata`). To restrict a skill to some agents, set `metadata.agents: claude openclaw hermes` (space-separated subset).
6. **Validate and wire up** from the repo root:

   ```sh
   python3 scripts/validate.py && python3 scripts/catalog.py
   scripts/link.sh
   ```

7. **Commit.** Never put secrets in skill text — skill files are injected into model context; reference credential file paths or env var names instead.