Claude Skills Examples: Twelve Skills Worth Copying

The best Claude skills examples to learn from are the ones Anthropic publishes in its own repository: a skill that builds skills, a document co-authoring workflow, an internal-comms writer, an MCP server builder, a web-app tester, a design guide, a brand guide and the PDF skill behind Claude’s file creation. Below are eight of those, named as they appear in the repository, and four of our own, marked as our examples. For each: what it does, the lines of its SKILL.md that carry the weight, and the request that makes it trigger. Building one from an empty folder is covered in how to create a Claude skill, and moving skills between Claude, Codex and Cursor in agent skills.

Eight from Anthropic’s repository

Anthropic’s skills repository↗ (opens in a new tab) keeps each skill in its own folder under skills/. Most are Apache 2.0; the four document skills (docx, pdf, pptx, xlsx) are source-available rather than open source. The README describes them as provided for demonstration and educational purposes, and warns that what Claude does in the apps may differ from these copies. Quotes below are from the SKILL.md files, shortened.

1. skill-creator

Writes a skill with you, then tests it. The description reads “Create new skills, modify and improve existing skills, and measure skill performance”, and the body turns that into a loop: draft the skill, “create a few test prompts and run claude-with-access-to-the-skill on them”, then review the results with you. It triggers when you ask to make, edit or benchmark a skill. Worth reading first, because it shows the evaluation habit the other eleven should have.

2. doc-coauthoring

A guided workflow for specs, proposals and decision docs in three stages: “Context Gathering, Refinement & Structure, and Reader Testing”. The description carries its own trigger list: “Trigger when user mentions writing docs, creating proposals, drafting specs”, and the body adds “PRD”, “design doc”, “RFC”. Notice that it offers the workflow rather than forcing it, which is the right manner for a skill that takes over a long conversation.

3. internal-comms

Status reports, leadership updates, newsletters and incident reports in a house format. The useful pattern is in the body: identify the kind of message, then load one file from examples/, such as examples/3p-updates.md for Progress, Plans, Problems. Only the file for the job enters the conversation. It triggers on any request to write internal communications. Its description says “help me write”, which Anthropic’s own guidance would put in the third person.

4. mcp-builder

A guide to building an MCP server, from research and planning to evaluation. The description names both the job and the stacks: “Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).” Naming the languages is what lets it trigger on “write me a FastMCP server” without the letters MCP ever meaning anything else.

5. webapp-testing

Tests a local web app with Playwright, using bundled helpers such as scripts/with_server.py. The key line is an instruction about context: “Always run scripts with --help first” and do not read their source unless you must, because “these scripts can be very large”. Scripts are run as black boxes, so only their output costs anything. It triggers on requests to test, debug or screenshot a local front end.

6. frontend-design

A point of view on visual design, not a procedure. It opens by casting Claude as a design lead whose client “has already rejected proposals that felt cliché or templated”, then asks it to ground the design in the subject before choosing type and colour. It triggers when you build or reshape a UI. A good model for a skill that changes judgement rather than steps.

7. brand-guidelines

Anthropic’s own colours and typography as data: hex values such as #d97757 for the primary accent, and “Headings: Poppins (with Arial fallback)”. It triggers “when brand colors or style guidelines, visual formatting, or company design standards apply”. Swap in your palette and fonts and it is the simplest useful skill you can write: no scripts, just facts Claude cannot guess.

8. pdf

One of the four document skills that, according to the repository, power Claude’s file creation. Its trigger is blunt and broad: “If the user mentions a .pdf file or asks to produce one, use this skill.” The body is a quick start with pypdf, and form filling lives in a separate file: “read FORMS.md and follow its instructions”. That split is progressive disclosure in practice.

Four of our own

These are our examples, not Anthropic’s. Each is short on purpose, and each covers a job the eight above do not. They use only name and description in the frontmatter, so they upload to the Claude apps as well as working in Claude Code.

9. pr-description (our example)

---
name: pr-description
description: Writes a pull request description from the branch's commits and diff.
  Use when the user asks for a PR description, PR summary or what to put in a PR.
---
1. Run `git log main..HEAD --oneline` and `git diff main...HEAD --stat`.
2. Write: Why (one paragraph), What changed (bullets a reviewer can check),
   How it was tested (only what was actually run), Risks.
3. Never claim a test that is not in the history or the conversation.

Triggers on “write the PR description”. Rule 3 is the line that earns its place: it stops the description inventing a test run.

10. sql-review (our example)

Reviews a query against your schema. The body says “Read reference/schema.md for table and column names; never guess a column” and “flag any UPDATE or DELETE without a WHERE clause”. The schema file can be long because it is read only when a query is under review. Triggers on “check this SQL” or “why is this query slow”.

11. release-checklist (our example)

---
name: release-checklist
description: Runs the pre-release checklist for this repository.
disable-model-invocation: true
---
Copy this checklist and tick each item as you go:
- [ ] Tests pass on main
- [ ] Changelog entry exists for every user-visible change
- [ ] Version bumped
Stop and report at the first failure. Do not tag or publish.

disable-model-invocation is a Claude Code field, so this one runs only when you type /release-checklist, never because Claude thinks a release looks near. Leave that field out if you upload it to the apps, which accept only the standard fields.

12. file-bug (our example)

Turns a bug found in conversation into a task on your board. The body: “Search first with fenbs:fenbs_search. If an open task matches, comment on it with fenbs:fenbs_comment. Otherwise create one with fenbs:fenbs_create_item, kind bug, with steps to reproduce, expected and actual.” It triggers on “log this as a bug” or “file that”. More on why this lives on a board rather than in the skill is below.

What the good ones have in common

Read the twelve side by side and the patterns match Anthropic’s skill authoring best practices↗ (opens in a new tab):

• The description says what and when, with the words people type: file types, commands, “PRD”, “FastMCP”. Claude chooses a skill from its description alone, so vague ones such as “helps with documents” never fire.
• Descriptions are written in the third person, because they are placed in the system prompt. “Writes a PR description”, not “I can help you”.
• The body assumes Claude is capable. Explain only what it cannot know: your schema, your palette, your rule about tests.
• Detail sits in files one level below SKILL.md, which stays under 500 lines. internal-comms and pdf both work this way.
• Freedom matches fragility. Design advice can be loose; a release checklist or a migration should be exact steps.
• MCP tools are named with their server, as in fenbs:fenbs_search, so Claude finds the right one when several servers are connected.
• Evaluations come before polish: the guide suggests three test scenarios and a baseline without the skill. Test with each model you plan to use.

Where to find more, and what to check first

• In Claude Code: run /plugin marketplace add anthropics/skills, then /plugin install example-skills@anthropic-agent-skills or document-skills@anthropic-agent-skills.
• In the Claude apps: the help centre’s page on what skills are↗ (opens in a new tab) points to the directory under Customize, Skills, then + and Browse skills, which includes partner skills from companies such as Notion, Figma and Atlassian.
• In your organisation: on Team and Enterprise plans, colleagues can share skills with you or publish them to the organisation’s directory.
• On GitHub: lists titled “awesome Claude skills” collect community skills. Treat them as leads, not as vetted software.

A skill is instructions plus, often, code that Claude will run. Anthropic’s guide to using skills in Claude↗ (opens in a new tab) says to install skills only from trusted sources and to review one from a less-trusted source, including one shared by a colleague, before enabling it. Read every file, not just SKILL.md. A skill that tells Claude to fetch a web page or read an inbox is also a route for indirect prompt injection; the wider picture is in MCP security risks.

Skills that write to a board

Example 12 is the pattern we would copy most. A skill should hold how a job is done, never the list of jobs: a to-do list inside a skill is stale by the next day and invisible to everyone else. With fenbs connected over MCP, the skill files the bug on a board where each task is a feature, an enhancement or a bug, sitting in To Do, Next Up, In Progress or Completed. The connection holds your role on that board, narrowed by the scopes you approved (read, write, comment), so a skill that tries to move a task through a read-only connection is refused, whatever its instructions say. Every change is recorded under the assistant’s name. fenbs has no sprints, due dates or assignee field, so a skill written for a heavier tracker needs those steps taken out.