After ten chapters of theory, here's the canon: six real skills shipped today, ranged from the smallest viable one to the largest orchestrator. Read them like recipes — each teaches a different lesson about the form.
You've seen the theory: descriptions are triggers, SKILL.md should be short, supporting files do the heavy lifting. Now read the actual skills. Six of them, picked to span the form: from a fifty-line one-trick pony to a multi-file orchestrator that spawns agents.
For each one we'll write a short paragraph, then call out the lesson it teaches that the others don't. At the end we'll name the five patterns visible in all six.
Each skill below is real — they all ship in the user's ~/.claude/skills/ today. They're ordered by complexity, smallest first.
Frontmatter, a tiny body, one Python script. About fifty lines total. Takes a YouTube URL, returns the transcript.
The youtube-transcript skill is the minimum viable one worth studying. Two trigger fields in the frontmatter (description names "transcript", "subtitles", "captions"), four headings in the body (Usage / Defaults / URL Formats / Output), and exactly one supporting file (scripts/get_transcript.py). Nothing more. It does one thing — pull a transcript — and tells Claude how to format the result.
A workflow skill that runs your app, observes what it does, and reports back. Three-step recipe, bounded scope.
Where youtube-transcript wraps a single command, verify wraps a small chain: do something (launch the app), watch what happened (read output, screenshot, observe behaviour), summarise the result for the user. The body lays out the steps in order — but the skill stays small because each step is a named primitive, not a paragraph of prose.
A config helper that touches settings.json. Adds permissions, env vars, hooks. Defensive about state.
The interesting thing about update-config is what it does before editing anything: it identifies which settings file to touch (global vs project), checks whether the change already exists, and respects locality (project-level changes should not bleed into user settings). The body is heavier than the previous two skills because the carefulness has to be encoded — there's no shortcut for "be safe with shared state".
Deep instructions for building, debugging, and migrating Claude API apps. ~300 lines plus reference files.
The claude-api skill is the long body — it genuinely needs every line. It has to explain prompt-caching break-points, model-version migration paths (Opus 4.5 → 4.6 → 4.7), how to handle batch / files / citations / memory APIs, plus a precise SKIP clause so it doesn't fire on OpenAI codebases. The description alone names the trigger files (imports anthropic) and the disqualifiers (imports openai). Heavyweight tables and migration matrices live in linked reference files.
Wraps a prompt or slash command to fire on a recurring interval. Polling, babysitting, self-paced loops.
Most skills wrap an action. The loop skill wraps a timing pattern: take any prompt and run it again every five minutes, or until a condition holds, or self-paced. The trigger phrases tell you what it's for ("check the deploy every five minutes", "keep running /babysit-prs") and the body explains the variants. It also tells Claude when not to use it — single-shot tasks should just run.
An orchestrator. Reads templates, dispatches sub-agents, verifies output with Playwright. The largest skill in the user's setup.
The book-builder skill is the orchestrator pattern at full size. Its SKILL.md describes three build paths (single chapter, multi-chapter, add-to-existing), names the canonical reference file every agent must read, and routes the actual heavy work to templates/build-chapter.md and references/. The skill itself is mostly routing — it tells Claude what to read, what to dispatch, what to verify. The recipe lives in supporting files.
Six skills, ranged from fifty lines to three hundred, single-script to multi-agent orchestrator. They look different on the surface. Underneath, they all share five things.
claude-api skill literally says "SKIP when code imports openai").verify doesn't also write reports. update-config doesn't also run hooks. book-builder doesn't also do code reviews. If you'd want it as a separate verb, it's a separate skill.If a skill of yours doesn't have all five, it's not necessarily broken — but it's probably the reason something feels off when you invoke it.
Click any of the six cards to expand a detail panel: actual description excerpt, folder tree, when to study it, and the three lessons it teaches. Default state is collapsed — pick the one that looks closest to what you're trying to build.
Type what you want to build. The matcher compares your description against the six gallery skills' lessons and suggests which one or two are closest in shape. Try the example chips first to see how it picks.
Pick two or three skills to compare side-by-side: their frontmatter description, body length in lines, the supporting files they ship with, and the one-line "signature pattern" each one represents.
The first tab is an excerpt of the book-builder SKILL.md — the largest orchestrator in the canon. The second is the full text of youtube-transcript — the smallest viable one. Read them side by side to see the spread.
The shapes are different — one routes work to templates, the other defines the work inline — but the patterns are the same. Specific triggers, clear "when to use" boundary, scannable body, one job each.