Top 10 Claude Code skills, ranked by GitHub stars

A skill is a folder holding a SKILL.md file (plus scripts and docs when needed) that Claude Code loads when a task matches its description. The ten sets below are the ten skill repos with the most stars at the time of measurement, with the install command for Claude Code and the places where they step on each other. None of them is required: Claude Code runs fine with nothing installed, and installing all ten costs context while the jobs overlap.

The ranking: 10 skill repos with the most stars

Stars were measured through the GitHub API on 29 Oct 2026. The top two are only 1,758 stars apart, so first and second place could swap within a few days.

Horizontal bar chart of GitHub stars for ten skill repos, Superpowers first with 307,991 stars, Taste Skill tenth with 105,112 stars

Stars measure popularity. They say nothing about how well a set fits your project. The ten sets are also not the same kind of thing: some are workflows, some are rule files, some are tools. The table below groups them by the job you need done.

Job	Sets to look at first
A workflow from idea to merge	Superpowers, Matt Pocock Skills, Agent Skills
Stop the agent from overdoing it and bloating the code	Karpathy Skills, Ponytail
Create Word, PDF, PowerPoint and Excel documents	Anthropic Skills
A UI that looks less like a stock template	UI UX Pro Max, Taste Skill
Ask the codebase through a graph instead of grep	Graphify
Cut the number of tokens the agent writes	Caveman

1. Superpowers: a development workflow with skills that trigger themselves

Superpowers (MIT) is a software development method packaged as skills. The agent does not rush into writing code: it asks what you want to build, presents the design in chunks short enough to read, then writes a plan of 2 to 5 minute tasks, each with file paths, complete code and a verification step.

The default chain, per the README: brainstorming, using-git-worktrees, writing-plans, subagent-driven-development or executing-plans, test-driven-development, requesting-code-review, finishing-a-development-branch. The TDD skill enforces the red, green, refactor cycle and deletes code written before its test. Because the skills trigger themselves, you do not type any command.

bash

/plugin install superpowers@claude-plugins-official

This repo runs the brainstorming, writing-plans, executing-plans chain, which is why every large job has a plan file before it has a commit. The cost is a few extra rounds of questions before you see the first line of code, so for a one-line fix the README lets you agree on it right in the chat.

2. Matt Pocock Skills: a round of questions before any code

Skills for Real Engineers (MIT) is the skill set Matt Pocock uses every day. He positions it as small, composable skills that work with any model, and avoids the kind of process framework that takes control away from you.

The best-known skills are /grill-me (for work that is not code) and /grill-with-docs: the agent asks you detailed questions, one at a time, about what you plan to do before it touches the code. The second part is a shared-language document for the project, so the agent stops using 20 words where 1 would do.

bash

claude plugin install mattpocock-skills@claude-plugins-official

Then run /setup-matt-pocock-skills once per repo. It asks which issue tracker you use, how you label issues and where you keep docs.

3. Karpathy Skills: one CLAUDE.md file, four principles

andrej-karpathy-skills is a single CLAUDE.md file, drawn from the mistakes Andrej Karpathy sees LLMs make when they write code: picking one reading of the task and running with it, writing 1,000 lines when 100 would do, and editing comments and code that have nothing to do with the task.

The four principles: think before coding (state assumptions, ask when something is unclear), simplicity first (no features beyond the request), surgical changes (touch only what must change), and work toward goals you can verify. This is the lightest set on the list because it is only text. To use it, download it as the CLAUDE.md of a new project:

bash

curl -o CLAUDE.md https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/main/CLAUDE.md

If the project already has a CLAUDE.md, append to the end instead of overwriting it. GitHub does not detect any license for this repo, so read it carefully before copying it into a commercial product. How to write a CLAUDE.md of your own is covered in writing a proper CLAUDE.md.

4. Anthropic Skills: the official sample library and document skills

anthropics/skills is Anthropic's own skill repo. It holds skills for creating docx, pdf, pptx and xlsx files (the source is viewable but not open source), and many sample skills under Apache 2.0 such as skill-creator, mcp-builder, webapp-testing, frontend-design and claude-api.

The repo states that these skills are for demonstration and learning, that the behavior you get in Claude may differ, and that you should test them yourself. The sensible way to use them is as templates for writing your own skills.

bash

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

5. Ponytail: make the agent write less code

Ponytail (MIT) gives the agent the mindset of "the laziest senior dev in the room": the best code is the code you never write. Its rule: write only what the task needs, and never cut input validation, error handling, security or accessibility.

The author's numbers for Ponytail 5: 53% less code, 41% less time, 26% lower cost, 45% fewer tokens, while 98% of risky logic has tests (68% without the skill). Measured in Claude Code with Opus 5.5, 39 tasks, 5 runs per task. The repo published these numbers itself, and I have not rerun them.

bash

/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail

Ponytail and Karpathy Skills both target code bloat, so pick one.

6. UI UX Pro Max: a lookup library for UI design

UI UX Pro Max (MIT) is a skill that gives Claude Code UI/UX design rules to look up: styles, palettes, font pairings, UX guidelines and chart types, per stack. The lookup runs as a Python 3 script, so you need Python installed.

bash

/plugin marketplace add nextlevelbuilder/ui-ux-pro-max-skill
/plugin install ui-ux-pro-max@ui-ux-pro-max-skill

My repo has it on through frontend-kit. It pays off when you ask "which style, colors and fonts should a dashboard for industry X use" and get an answer grounded in the lookup tables instead of the model's random taste.

7. Graphify: ask the codebase through a knowledge graph

Graphify (Apache-2.0) turns a codebase, along with its docs, SQL schemas, config files and PDFs, into a knowledge graph you can query. Code is parsed locally with tree-sitter, deterministically, without an LLM; only docs, PDFs and images need a model for the semantic step. Every edge in the graph is labeled EXTRACTED (present in the source) or INFERRED (deduced by Graphify), so you know which ones were read directly and which were guessed. There are no embeddings and no vector store.

bash

uv tool install graphifyy
graphify install

Then type /graphify . in your AI assistant. You get a graph.html that opens in a browser, a GRAPH_REPORT.md summarizing the main concepts, and a graph.json for querying again without rereading the files. It needs Python 3.10 or later. The repo also sells a Graphify Cloud service (14-day trial), which is separate from the free skill. codebase-memory-mcp follows the same idea of querying a graph instead of grepping; pick one of the two per repo.

8. Caveman: make the agent talk less to save tokens

Caveman (Apache-2.0) makes the agent answer in clipped caveman speech. The README compares the same answer about a React re-render bug: 63 tokens normally, 20 tokens with /caveman, 14 with /ultracave (counted with tiktoken o200k). The repo claims a 65% token cut.

bash

claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman

My repo runs it at the ultracave level. The price is curt answers, so turn it off when you want a long explanation or when the reader is someone other than you. The 65% figure comes from the repo; your real savings depend on whether the agent's current job is mostly talking or mostly reading.

9. Agent Skills by Addy Osmani: 25 skills across the engineering lifecycle

agent-skills (MIT) packages a senior dev's workflow into 25 skills and 9 commands, /spec, /plan, /build, /test, /constraints, /review, /webperf, /code-simplify, /ship, which map to the steps define, plan, code, test, review, ship. The skills also trigger themselves based on what you are doing: designing an API loads api-and-interface-design, building a UI loads frontend-ui-engineering.

bash

npx skills add addyosmani/agent-skills

Or use the plugin: /plugin marketplace add addyosmani/agent-skills. In Claude Code, /review clashes with a built-in command, so call this set's version as /agent-skills:review. This set covers almost the same lifecycle as Superpowers, so do not turn both on in one repo.

10. Taste Skill: force the agent to have design taste

Taste Skill (MIT) targets web pages that "look like an AI template": the agent follows an art direction, works from reference images and checks the result in a browser, instead of producing a generic page.

bash

/plugin marketplace add Leonxlnx/taste-skill
/plugin install taste-skill@taste-skill

If you do not use the plugin, the README gives the command npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend". Taste Skill and UI UX Pro Max both target the UI: one works from taste and reference images, the other looks up a rule table. Try both on one page and keep the one whose result you like.

Stacking several sets: three overlaps where you pick one

Workflow: Superpowers, Matt Pocock Skills and Agent Skills all cover asking clarifying questions, planning, TDD and review. Matt Pocock's README also warns that installing the same set two ways (plugin and skills.sh) gives you every skill twice. Keep one workflow set per repo.
Fighting code bloat: Karpathy Skills and Ponytail.
UI: UI UX Pro Max and Taste Skill, plus frontend-design in Anthropic Skills.

Caveman, Graphify and Anthropic Skills add to a setup rather than replacing any other set, so they combine with a workflow set.

Skills are instructions the agent actually runs, so read them before installing

A skill is more than text to read: it can ship scripts and tell the agent to run commands on your machine. The Anthropic Skills repo says plainly that you must test thoroughly in your own environment, and that holds for the other nine sets too. Open SKILL.md and the bundled scripts before installing, pin a version if the repo has releases, and do not install a set just because it ranks high on the star table.

Limits of this ranking

GitHub stars reflect how far a repo spread, including the times it reached the Hacker News front page or trending; they say nothing about quality. This table will also change: the stars were measured on 29 Oct 2026, and the top two are close. I have only run Superpowers, UI UX Pro Max and Caveman for real in this repo; for the other seven I read only the README and the public source, so their descriptions rest on the official docs and have not been through a project of mine. The safest approach: pick one workflow set and one set for the problem you hit most often, run them for two weeks, then add more.