Skill Publishing

skill-publishing

How to write a good agent skill and publish it so others (and your other machines) can install it with npx skills.

What a skill is

Reusable, on-demand context for an AI coding agent. One skill = one folder with a SKILL.md:

---
name: my-skill            # kebab-case, matches the folder
description:     # the TRIGGER — see below; most important field
metadata:
  author: you
  version: "1.0.0"        # semver; bump on meaningful change
---

# Title

Context economy: only the description of every installed skill stays in the agent's context. The body loads only when the description matches the task. That's the whole design — so the description does the heavy lifting.

The description is the trigger (get this right)

Write it as concrete triggers: what it does, then "Use when …" listing symptoms, situations, and phrases a user might actually say.

• Good: `…Use when an iPhone page shows black bars, content is cut off at the bar

edge, a cropped shadow, or the page jumps after the keyboard closes.`

• Bad: Helps with iOS Safari styling.

If the agent won't match it to the right moment, the skill is dead weight.

The body

• Write for the agent, not end users: actionable facts, steps, rules.
• Scannable: headings, short sections, tables, code blocks.
• Progressive disclosure: keep SKILL.md focused; put heavy reference docs or

scripts in sibling files and link them so they load only when needed.

• Prefer facts + case-dependent guidance over one rigid recipe, so it stays

useful across situations.

• Note provenance for non-obvious claims (how it was verified) to earn trust.
• One concern per skill — split rather than overload.

Bundling files in a skill (progressive disclosure)

A skill is a directory, not just one file. SKILL.md is the entry point; you can ship reference docs, scripts, templates, and assets next to it:

skills/my-skill/
  SKILL.md                 # entry: frontmatter + concise body
  reference/
    api.md                 # deep detail / long tables — read on demand
  scripts/
    init.sh                # a runnable helper the skill invokes
  templates/
    config.example.json    # boilerplate the skill copies

Only SKILL.md's body loads when the skill fires; everything else loads, reads, or runs only when SKILL.md tells the agent to, referenced by relative path. Keep SKILL.md short and high-signal; push heavy material into siblings:

For the full option list, read `reference/api.md`.
Scaffold with `scripts/init.sh `.

The CLI installs the whole directory (symlink or --copy), so relative links keep working. Split only when a skill gets large or needs runnable helpers — a small skill is perfectly fine as a single SKILL.md.

Repo layout (bundling many skills)

skills//SKILL.md            # flat (preferred)
skills///SKILL.md # categorised (only if you have enough)
SKILL.md                                # a single skill at repo root also works

• No manifest required — the CLI auto-discovers those paths.
• Don't put a root SKILL.md alongside a skills/ dir: a root SKILL.md

short-circuits discovery — the CLI treats the repo as one skill and never scans subdirectories unless consumers pass --full-depth. Use one layout or the other.

• README.md = the human index (a table of skills) + SEO. The machine index

is just each skill's description; no separate index file.

• AGENTS.md (+ CLAUDE.md symlink) = authoring conventions for the repo.
• A .claude-plugin/marketplace.json manifest is optional, only for

plugin-marketplace features or non-standard skill paths.

The skills CLI (vercel-labs/skills)

npx skills init skills/        # scaffold a new skill folder + SKILL.md

npx skills add owner/repo            # interactive: pick skills + agents
npx skills add owner/repo --list     # browse the repo, install nothing
npx skills add owner/repo --skill a b   # install specific skills
npx skills add owner/repo --all -g   # everything, all agents, global — no prompts
npx skills add owner/repo --copy     # copy instead of symlink into agent dirs

npx skills list                      # installed skills (-g for global)
npx skills find             # search
npx skills update [name...]          # update
npx skills remove [name...]          # remove
npx skills use owner/repo@skill      # one-off prompt without installing

Flags that matter: -s/--skill, -a/--agent (* = all), -g/--global, -l/--list, -y/--yes (skip prompts), --all, --full-depth (scan every subdirectory even when a root SKILL.md exists — needed for a repo that has both a root skill and a skills/ dir). Telemetry is on by default; set DISABLE_TELEMETRY=1 (or DO_NOT_TRACK=1) to opt out.

> Provenance: CLI commands, flags, and behaviour verified against skills > v1.5.x (2026-07). The CLI iterates fast — re-check flags if it's since moved.

Publishing to GitHub

git init && git add -A && git commit -m "chore: bootstrap skills repo"
gh repo create owner/skills --public --source=. --push
# SEO: a clear description + topics make it discoverable
gh repo edit owner/skills \
  --description "Agent skills for … — install: npx skills add owner/skills" \
  --add-topic agent-skills --add-topic claude-skills --add-topic claude-code \
  --add-topic skills --add-topic ai-agents

Keep the README's skills table and each description sharp — those are what both humans and npx skills find read.

Public vs private

npx skills add git clones the repo under the hood (defaulting to the HTTPS URL) — apart from a no-clone GitHub-API fast path reserved for a few allowlisted first-party owners (e.g. vercel), any other repo is cloned. So:

• Public → clones with no auth.
• Private → works too, if the machine is authed to clone it. For HTTPS,

run gh auth setup-git once so git uses your GitHub login non-interactively; or pass an SSH URL directly: npx skills add git@github.com:owner/repo.git.

• Anyone you share the install command with also needs repo access + auth, and

public discovery surfaces (skills.sh leaderboard/search) only index public repos. So: private = great for personal/studio use; public = shareable + discoverable.

Checklist before publishing a skill

• [ ] Folder is skills// and name: matches.
• [ ] description reads as triggers ("Use when …"), not marketing.
• [ ] Body is actionable, scannable, one concern.
• [ ] metadata.version set (semver).
• [ ] Added to the README skills table.
• [ ] npx skills add --list shows it with the right description.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: stealth-engine
• Source: stealth-engine/skills
• License: MIT

Install and usage instructions live in the source repository linked above.