Claude Code Skills Examples: 7 Real Skills and What Their SKILL.md Files Actually Do

Claude Code skills are markdown instruction files that teach the agent a specific job: how to design a UI, how to pressure-test a plan, how to write Postgres queries, how to drive a browser. The fastest way to understand them is to read real ones. Below are seven skills from the live catalog, each covering a different job type, with what their SKILL.md actually instructs and when you would reach for each.

Key takeaways

• A skill is a SKILL.md file: YAML frontmatter (name + description) plus plain-markdown instructions, optionally bundled with scripts and reference files.
• Real skills span very different jobs: discovery, UI generation, plan review, process enforcement, framework rules, browser automation, and database tuning.
• The most-installed example, find-skills, has 2,438,642 installs (skills.sh registry, July 2026) - and it is itself just a markdown file describing a six-step workflow.
• The description field in the frontmatter decides when Claude auto-loads the skill; the body decides what it does once loaded.
• You can copy any pattern here into your own skill - the format is identical whether the skill has 100 installs or 2 million.

What do real Claude Code skills look like?

Seven examples, each a distinct job type. Every install count is from the skills.sh registry, July 2026.

1. find-skills - meta and discovery

find-skills by vercel-labs, 2,438,642 installs. Its SKILL.md instructs the agent to run a six-step discovery flow when you ask "how do I do X": check the skills.sh leaderboard for proven options, search with npx skills find, verify install counts and source reputation before recommending, then present install commands. It also encodes quality gates - prefer skills with 1K+ installs and trusted publishers. Reach for it when you suspect a skill already exists for your task and do not want to reinvent it.

2. frontend-design - UI generation

frontend-design by anthropics, 649,620 installs. The SKILL.md forces the agent to commit to a bold aesthetic direction (purpose, tone, one differentiation point) before writing any UI code, and explicitly bans generic-AI tells: overused fonts like Inter, purple-gradient-on-white cliches, cookie-cutter layouts. Reach for it when Claude's default UI output looks like every other AI-generated page.

3. grill-me - plan pressure-testing

grill-me by mattpocock, 515,987 installs. Its instructions run an interview: the agent asks questions about your plan one at a time, each with its own recommended answer, and waits for your feedback before continuing. Facts it can find in the codebase it looks up itself; only genuine decisions go to you, and it will not start implementing until you confirm shared understanding. Reach for it before any non-trivial build, while wrong assumptions are still cheap.

4. test-driven-development - process enforcement

test-driven-development from obra's superpowers collection, 160,840 installs. The SKILL.md enforces RED-GREEN-REFACTOR: no production code without a failing test first, verify the test actually fails before implementing, and treat the agent's own rationalizations for skipping tests as red flags. Reach for it when you want the discipline applied to every feature and bugfix, not just when you remember to ask.

5. vercel-react-best-practices - framework rules

vercel-react-best-practices by vercel-labs, 542,213 installs. It embeds 70 React and Next.js performance rules across 8 impact-ranked categories - eliminating network waterfalls, bundle size, server-side performance, re-render optimization - and applies them whenever the agent writes or reviews components. Reach for it during performance audits and PR reviews where generic React advice is not specific enough.

6. agent-browser - browser automation

agent-browser by vercel-labs, 533,255 installs. This one shows skills can be thin wrappers around real tools: the SKILL.md is a discovery stub that tells the agent to load full workflows from a CLI (agent-browser skills get core), which then drives Chrome over CDP using accessibility-tree snapshots and compact element refs. Reach for it when Claude needs to log in somewhere, fill forms, screenshot pages, or automate an Electron app.

7. supabase-postgres-best-practices - data and API work

supabase-postgres-best-practices by supabase, 280,017 installs. The SKILL.md organizes Postgres performance guidance into 8 priority categories with incorrect-vs-correct SQL examples, query plans, and rules for indexes, RLS, and concurrency. Reach for it whenever the agent writes queries or schema against Postgres, so it applies vendor-maintained rules instead of remembering generic SQL advice.

Browse the full catalog at /skills or see what developers actually install on the leaderboard.

What is the anatomy of a SKILL.md?

Every example above follows one format:

• Frontmatter: YAML with name and description. The description is the trigger - Claude reads it to decide when to load the skill, so it should state the task and the situations that call for it.
• Body: plain markdown instructions. Workflows, rules, checklists, banned patterns - whatever the agent should follow once loaded.
• Optional bundled files: scripts the agent can run and reference docs it can read on demand (superpowers and agent-browser both use this).

Here is a minimal skeleton - invented for illustration, not a published skill:

---
name: my-team-api-conventions
description: Use when writing or reviewing REST endpoints in this repo -
  enforces our error format, pagination style, and naming rules.
---

# API conventions

## When writing an endpoint
1. Errors always return { "error": { "code", "message" } }.
2. List endpoints paginate with ?cursor= and ?limit= (max 100).
3. Run scripts/check-openapi.sh after changing any route.

## Banned patterns
- No offset pagination.
- No 200 responses with an error body.

That is the whole format. Directory layout: .claude/skills/my-team-api-conventions/SKILL.md, with optional scripts/ and references/ alongside. For the full conceptual background, see what Claude Code skills are.

How do you go from example to your own skill?

Steal the shape, not the content. Pick the example closest to your job type - rules skill (vercel-react-best-practices), workflow skill (grill-me), tool wrapper (agent-browser) - write a sharp description so it triggers at the right moment, and keep the body to instructions you would give a new teammate. The step-by-step walkthrough, including testing that the skill actually triggers, is in how to build custom Claude Code skills. For a curated map of which community skills are worth studying first, see the Matt Pocock skills map.

Install counts come from the skills.sh registry via the Skillselion catalog, snapshot July 2026.

Common questions