Search skills/

Guide · 8 min read · updated Oct 29, 2026

Claude skills examples worth reading before you write your own

Claude skills examples are easiest to learn from when they are real: a SKILL.md that people actually install, with its scripts and references visible. Below are 21 skills from our catalog, grouped by the job they do. For each one we note what is in the folder, what it can touch on your machine, and the verdict from our review method. If you are new to the format, read what Claude skills are first.

A note on reading the labels. “Recommended” and “With caveats” are our verdicts on usefulness. “No issues found”, “Review before installing” and “Broad access” are separate security statuses based on the files we could read. Install counts come from the data we collected on 2026-10-29.

Document skills: Word, PowerPoint and brand rules

Document skills are the most official examples. Anthropic ships pre-built skills for PowerPoint, Excel, Word and PDF on claude.ai and the API,, and its skills repository on GitHub carries the docx and pptx skills as folders you can read. Anthropic’s docs say the pre-built document skills are not available in Claude Code.

  • docx creates and edits Word files, including tracked changes and comments. A 7 KB SKILL.md sits on about 15 Python scripts, XML templates and validators. It runs local helpers with LibreOffice and pandoc and makes no network calls. Recommended, Review before installing, proprietary license.
  • pptx builds decks with pptxgenjs, edits existing ones at the XML level, and requires a validation and rendering pass. A long list of gotchas (no # in hex colors, no shared option objects) is the real value. Recommended, Review before installing.
  • brand-guidelines is the opposite end: a 2 KB palette and font list with no scripts. It shows how small a skill can be and still be useful, as a pattern for your own brand file. With caveats, No issues found.

What to copy: a short SKILL.md that routes to scripts, plus a validator that checks the output. The docx and pptx skills do not trust the model to produce a valid file; they check it.

Planning and handoff skills

These skills change how a session runs rather than producing a file.

  • grill-me is the most installed single skill in our catalog, at about 1.45 million installs, and its SKILL.md is a single instruction: call the grilling skill. It sets disable-model-invocation, so it only runs when you type /grill-me. The behavior lives in grilling, which interviews you about a plan in rounds until no open branch remains. With caveats, because installing it alone does nothing.
  • handoff writes a summary of the current conversation to the system temp directory so a fresh agent session can continue. It is Markdown only and tells the agent to redact secrets, but the model does that redaction, so check the file before you share it. Recommended, No issues found.
  • context-engineering is guidance on what to put in front of an agent: rules files, spec excerpts, error output. A single 15 KB file with checklists. Recommended, No issues found.

What to copy: the pointer pattern in grill-me keeps a user-invoked entry point short while the shared logic lives elsewhere. Its cost is a hidden dependency, so say so in the README.

Coding skills

Coding is the largest category in the catalog, with more than a hundred entries tagged for dev workflow. Four examples with different shapes:

  • tdd is a test-driven development guide with two supporting documents on tests and mocking. It has no scripts or network calls, and the agent only writes tests at seams you agreed with it. Recommended, No issues found.
  • diagnosing-bugs is a six-phase debugging procedure that insists on a fast failing signal first and ranked hypotheses shown to you before any testing. It includes one template shell script that only asks you questions. Recommended.
  • code-review reviews the diff since a fixed point with two parallel sub-agents, one for standards and one for the spec. It uses read-only git commands and may call gh or glab to fetch the linked issue, which is the reason for its Review before installing status. Recommended.
  • vercel-react-best-practices packs 70 React and Next.js performance rules into separate files with incorrect and correct code examples, plus an index. The agent loads only the rules a task needs, which is progressive disclosure at its clearest. Recommended, No issues found.

What to copy: the Vercel layout. A small index in SKILL.md, one file per rule, one wrong and one right example in each. It scales to dozens of rules without a large context cost.

Design and front-end skills

  • frontend-design is a guidance-only brief for avoiding template-looking UI: palette, type, layout and UI writing. One 9.4 KB file and a license, over a million installs. Recommended, No issues found.
  • web-design-guidelines is a 1.2 KB file that fetches Vercel’s interface guidelines from a GitHub branch every time it runs and then reports file:line findings. The rules can change without the skill changing. With caveats, Broad access, and a good lesson in why we flag remote instructions.

Compare the two. Both are useful design aids and both are short. One is fully inspectable; the other is a pointer to text you did not review.

Cloud and backend skills that pair with MCP and CLIs

Many vendor skills exist to give an agent the rules for a tool it can already call. This is the practical answer to “Claude skills MCP”: the MCP server or CLI supplies the capability, and the skill supplies the do-and-don’t list.

  • supabase tells the agent to check current docs, avoid common Auth and row-level security mistakes, and use the Supabase CLI and MCP server with scoped tokens. A 14 KB file with a changelog and references, no scripts. Recommended, Review before installing because it fetches docs at run time and uses your project credentials.
  • azure-deploy runs azd up, terraform apply or az deployment for a prepared app. About 40 files include recipes, four helper scripts and pipeline examples. It creates cloud resources, grants SQL roles and applies migrations, and one reference doc contains a curl | sh pattern. Recommended, Review before installing.
  • azure-resource-visualizer reads a resource group with read-only queries and writes a Mermaid diagram, with an instruction to show placeholders instead of secret values. Recommended, Review before installing.

Same vendor, same repository, very different blast radius: one skill reads, another changes live infrastructure. Check what a skill is allowed to do, not only which tool it uses.

Marketing and SEO skills

  • seo-audit is a Markdown checklist with seven reference files and 13 eval prompts. It has no scripts, so you provide any Search Console or browser access, and it tells the agent to treat fetched pages as untrusted data. Recommended, No issues found.
  • copywriting diagnoses your input before it drafts, and applies strict rules against AI-sounding phrasing and unsupported claims. Recommended, No issues found.

The eval prompts in seo-audit are the detail to copy. They let you test whether a change to the skill made it better.

Guardrail and style skills

  • git-guardrails-claude-code copies a script into .claude/hooks and adds a PreToolUse hook that blocks git push, git reset --hard and similar commands. It is a good example of a skill acting through a hook rather than through advice. Matching is plain text and it needs jq, and it edits your settings.json. With caveats, Review before installing.
  • caveman is a 4.7 KB output-style prompt that makes the agent answer tersely until you say stop. The skill folder holds only SKILL.md and display metadata. Recommended; its status is Review before installing only because we could not download two of the four files in the large repository.

Skills for writing skills

  • skill-creator is Anthropic’s skill for drafting a skill, running test prompts with and without it, and tuning the trigger description. A 33 KB SKILL.md with 10 Python scripts and a review viewer, depth 5. The scripts call claude -p and the viewer loads fonts and a pinned script, so run it in a project you trust. Recommended, Review before installing.
  • writing-for-agents is a reference on writing skills, AGENTS.md and CLAUDE.md: context pointers, progressive disclosure, completion criteria. Recommended, No issues found.

A Claude skills template to start from

The patterns above reduce to a small template. Replace the text and keep the structure:

---
name: your-skill-name
description: Claude skills examples from real repositories, grouped by job: documents, coding, design, cloud and SEO. See what each does, what it can access, our verdict.
---

# Your skill name

## When to use this
One or two lines on the trigger and the cases to skip.

## Steps
1. First action, with the exact command or file if there is one.
2. Second action.
3. Check the result before reporting it.

## Reference
For details, read [reference.md](reference.md) only when needed.

Add a scripts/ folder only for work that should be exact, a references/ folder for detail the agent reads on demand, and disable-model-invocation: true if the skill has side effects. The SKILL.md format guide explains each field.

CLAUDE.md vs skills: where each example would go

Several examples above could have been written as a CLAUDE.md section. The test is when you want the text loaded. CLAUDE.md is always in context, so it suits short rules that apply to every task: build commands, naming conventions, folders to avoid. A skill suits procedures you need sometimes, like a deploy runbook or a 70-rule performance guide, because its body loads only on use. Anthropic’s docs recommend moving procedures out of CLAUDE.md into skills. See skills vs subagents and skills vs MCP for the other boundaries.

What these examples show

Three patterns repeat across the skills that are worth installing. They keep SKILL.md short and push detail into files read on demand. They check their own output with a validator, a test or an eval. And they tell you what they touch: secrets are redacted, hooks are offered rather than silently installed, and destructive steps ask first.

The pitfalls repeat as well: pointers to text fetched at run time, skills that only work with a second skill installed, and tools that edit your configuration. Browse the full list of reviewed skills or filter by job on the category pages to find examples closer to your own work, and check our security notes before running anything with scripts.