Shipping Skills

Shipping Skills

Overview

This skill is the end-to-end playbook for taking an Agent Skill from idea to a published, discoverable GitHub repo. It is opinionated and battle-tested: every step here exists because skipping it caused a real, avoidable problem.

The pipeline: design → write SKILL.md → triggers + evals → repo structure → name + discoverability → cross-file consistency → publish to GitHub → distribution → pre-publish review.

When to Use / Not Use

Use when: creating a new skill, writing or fixing a SKILL.md, structuring a skill repo, naming it, publishing to GitHub with releases, or getting it discovered.

Do not use for: writing application code, non-skill documentation, or general Git questions unrelated to shipping a skill.

1. Design first

Before writing anything, answer three questions:

• What cognitive/work failure does it fix, in one sentence? A skill that does "many things" triggers reliably for none.
• When should it fire? This becomes the description. Write down 5 example user phrases that should trigger it and 5 that should NOT.
• Single skill or a repo of skills? Standalone repo or fold into an existing one? Decide by mechanism + coupling + maintenance bandwidth, not vibes. A tool whose main use is independent of an existing repo belongs in its own repo. Don't put a tool inside a repo whose name implies a different mechanism (it misleads users about how it works).

2. Write the SKILL.md

Required frontmatter: name and description. Add version, license, metadata.tags.

The description is the single most important field — agents load only name + description at startup and use it to decide whether to activate the skill (progressive disclosure). So:

• Lead with user-intent trigger phrases ("second opinion", "fact-check this", "review my plan"), NOT internal-mechanism jargon ("heterogeneous multi-agent orchestration"). Users don't speak in your architecture's terms.
• Include a negative boundary: "Do not use for …". This stops the skill from grabbing tasks that belong to other skills.
• Keep the body focused; push large reference material into references/ and load on demand (progressive disclosure). A bloated SKILL.md wastes every session's context.

Body essentials: Overview · When to use/not · the core protocol/steps · examples · a verification checklist.

3. Triggers + evals

Ship an evals/ set from day one: ~10 should-trigger and ~10 should-not-trigger prompts. The boundary cases (prompts that look like they fit but belong to a neighboring skill) are the highest-value ones — they stop misfires. Re-run after every description edit; it's a regression test.

4. Repo structure

Use the portable layout so the skills CLI discovers it:

my-skill/
├── skills//SKILL.md   # the skill (CLI scans skills//SKILL.md)
├── README.md                      # result-first; what it does, install, when to use
├── LICENSE                        # e.g. MIT
├── CHANGELOG.md
├── evals/                         # trigger eval set
├── assets/                        # banner, etc.
└── references/                    # optional deep docs (progressive disclosure)

5. Name + discoverability (a common trap)

• Check for name collisions before committing to a name. Generic names ("agent-arena", "fact-check") collide with big projects and dilute search. Pick something descriptive + distinctive that does not imply the wrong mechanism.
• Make the GitHub repo description carry category keywords + a user-intent phrase.
• Add topics: claude-code-skill, agent-skill, plus your domain — directories often auto-index by topic.

6. Cross-file consistency (a trap I hit)

If the skill spans multiple spec files (SKILL.md + interop/schema docs), declare ONE file the normative source for every enum/contract and have the others defer to it. Otherwise the files drift (e.g. code-api-behavior in one, code-api in another) and break integrators. Verify with a quick grep that enums/terms match across files before publishing.

7. Publish to GitHub

# local
git init -b main && git add -A && git commit -m "feat: initial  scaffold (v0.1.0)"

# create public repo and push (gh CLI)
gh repo create / --public --source=. --remote=origin --push \
  --description ""

# topics for discoverability
gh repo edit / --add-topic claude-code-skill --add-topic agent-skill

# tag + release = a trust signal for a new repo
git tag -a v0.1.0 -m "v0.1.0 — initial preview" && git push origin v0.1.0
gh release create v0.1.0 --title "v0.1.0 — " --notes ""

Then verify — two distinct checks, don't conflate them:

# 1. Frontmatter parses & the skill is discoverable (does NOT install):
npx skills add / --list   # must print "Found N skills" + your skill name
# 2. Smoke-test the real install path end-to-end:
npx skills add / -g -a claude-code -y

--list only confirms the frontmatter parses and the skill is listed — it is not proof the install works (and a failure isn't only a frontmatter problem). Run the actual add (#2) before announcing.

8. Distribution (how skills actually get found)

• skills.sh — no manual submit. The directory/leaderboard is built from the skills CLI's anonymous install telemetry. You appear by getting people to run npx skills add /. So: promote the install command (README, a launch post, socials).
• agentskills.io — this is the open-standard spec site, not a submission directory. You don't "publish" to it; you just conform to the standard (valid SKILL.md). Engage via its GitHub/Discord if you want.
• Marketplaces (e.g. mcpmarket) — submission varies by site: some auto-index public repos by topic, some have a submit page (mcpmarket has one). Check each site's own process; don't assume one path fits all.
• Highest-leverage single action: a one-line npx skills add command + a short launch post telling the result-first story. That feeds the leaderboard and lets auto-indexers find you.

9. Review before publishing

Run a heterogeneous review (e.g. a second model / agent) over the skill before announcing. It reliably catches what you can't see in your own work: cross-file enum drift, misleading wording, over-claims, and README commands that don't actually run. Fix, bump a patch version, then publish.

Common Mistakes (each one bit me)

1. Mechanism-jargon description — users never trigger it because they don't speak your architecture's language. Use intent phrases.
2. No negative boundary — the skill grabs tasks that belong elsewhere.
3. Unchecked name collision — buried under same-name giants in search.
4. Name implies the wrong mechanism — e.g. a single-agent tool in an "arena" repo misleads users about how it works.
5. Cross-file enum drift — multiple spec files with no normative source; integrators break.
6. Announcing without a real install smoke test — --list only checks frontmatter parsing, not that the install works; run npx skills add ... -g -a -y first.
7. Treating directories as manual submit forms — skills.sh is automatic via install telemetry; promote the install command instead.
8. No release/tag — not required, but a 0-release repo reads as abandoned; a tagged release is a cheap trust signal (nice-to-have, not a hard rule).

Verification Checklist

• [ ] description leads with user-intent phrases + has a "Do not use" boundary.
• [ ] evals/ has should-trigger AND should-not-trigger prompts, incl. boundary cases.
• [ ] One normative source declared for any cross-file enums/contracts; grep-confirmed consistent.
• [ ] Repo has README (result-first), LICENSE, CHANGELOG, topics, description.
• [ ] Name checked for collisions; doesn't imply the wrong mechanism.
• [ ] --list prints the skill (frontmatter parses) AND a real npx skills add ... -g -a -y smoke test succeeds.
• [ ] Tag + GitHub release published.
• [ ] Heterogeneous review done; README commands actually run.

Example Prompts

• "Help me turn this workflow into an Agent Skill and publish it to GitHub."
• "Write a SKILL.md for X and tune its description so it triggers reliably."
• "Review my skill repo before I publish — naming, frontmatter, consistency."
• "How do I get my skill listed on skills.sh?"

Source & license

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

• Author: zhjai
• Source: zhjai/shipping-skills
• License: MIT

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