The 10 Most-Starred Agent Skills on GitHub, and Why Stars Are Not Trust

We checked GitHub star counts and read each README. Here are the 10 most-starred agent skills, what they do, and why you should read a skill before you install it.

Published
· 5 min read
Category
Tech
Topics
agent-skillsgithubclaude-codeopen-sourceai-coding
한국어 요약
GitHub에서 별(star)을 가장 많이 받은 에이전트 스킬 10개를 2026년 11월 11일 03:13(UTC) 기준 검색 API로 집계하고, 10곳 모두 저장소에 SKILL.md가 실제로 있는지 README와 함께 확인했습니다. 1위는 Anthropic 공식 저장소(약 19.2만 개)이고, 2위 ponytail(약 18.8만 개)은 올해 6월 만들어졌는데도 바짝 따라붙었습니다. 10개 중 8개가 올해 만들어졌고, 디자인·문서·말투처럼 에이전트의 '결과물 품질'을 다듬는 스킬이 많다는 것이 우리 해석입니다. 다만 별 수는 인기일 뿐 안전이나 품질의 보증이 아니며, ponytail의 절감 수치처럼 프로젝트가 직접 밝힌 벤치마크는 독립 검증되지 않았습니다. Anthropic 문서는 스킬을 직접 만들었거나 Anthropic에서 받은 것만 쓰라고 권하므로, 1위를 제외한 9개는 설치 전에 SKILL.md와 스크립트를 직접 읽어야 합니다.

An agent skill is a folder of instructions that a coding agent reads when a task matches it. The folder always has a file called SKILL.md. It may also hold scripts and templates. Think of it as a recipe card you hand to a cook who already knows the kitchen.

Skills are now one of the most popular things to share on GitHub. We ranked the ten most-starred ones, read their README files, and checked that each repository really contains skills. This is a list of popularity, not of quality. We explain why that difference matters at the end.

How we built the list

On November 11, 2026, at 03:13 UTC, we asked the GitHub search API for the most-starred repositories under the topics agent-skills, claude-skills, claude-code-skills, agent-skill, and claude-code-plugins. A star is a bookmark. It counts how many accounts saved the project.

We then removed projects that are not skills:

  • Apps and platforms, such as paperclipai/paperclip, nexu-io/open-design, and career-ops-hq/career-ops.
  • Curated link lists, such as ComposioHQ/awesome-claude-skills (79,232 stars).
  • Memory and tool projects, such as thedotmack/claude-mem (115,879 stars).
  • REA, which has 450,319 stars but is a tool that connects agents to reverse-engineering software. We covered it here.

For the ten that remained, we cloned each repository and counted its SKILL.md files. All ten have at least one. We did not run or audit any skill.

The top 10

# Repository Stars Created SKILL.md files What it does
1 anthropics/skills 192,438 Sep 2025 20 Anthropic’s example skills, from design to document editing
2 DietrichGebert/ponytail 188,466 Jun 2026 12 Pushes the agent to write less, simpler code
3 addyosmani/agent-skills 115,249 Feb 2026 25 Engineering steps from spec to ship
4 Leonxlnx/taste-skill 113,046 Feb 2026 1 Helps agents build better-looking company websites
5 pbakaus/impeccable 107,681 Nov 2025 26 A design vocabulary, with commands like polish and audit
6 tt-a1i/archify 95,415 Apr 2026 5 Turns ideas or plans into interactive HTML visuals
7 Egonex-AI/Understand-Anything 91,914 Mar 2026 9 Turns a codebase or docs into a searchable knowledge graph
8 blader/humanizer 67,613 Jan 2026 1 Rewrites text to remove common signs of AI writing
9 cathrynlavery/diagram-design 66,350 Apr 2026 1 Draws 45 kinds of diagrams as one HTML and SVG file
10 ayghri/i-have-adhd 60,781 May 2026 2 Makes agent answers action-first, with numbered steps

Next in line were coreyhaines31/marketingskills (59,604) and K-Dense-AI/scientific-agent-skills (56,743). Stars change every hour, and the top two are only about 4,000 apart, so their order could flip soon.

What stands out

Most of the list is new. Eight of the ten were created in 2026. Only Anthropic’s repository and Impeccable are older.

Many skills polish the result rather than add power. Four of the ten are about design or visuals, and two more shape how the agent writes or answers. Only a few, such as Ponytail and Addy Osmani’s set, change how the agent works on code. This is our reading of the READMEs, not a measured finding.

Licenses are mostly open. Nine repositories show MIT or Apache-2.0 licenses in the GitHub data. Anthropic’s repository shows none, but its README says many skills in it are open source under Apache 2.0, including the document skills for Word, PDF, PowerPoint, and Excel.

Some numbers come from the authors. Ponytail’s README reports that, in its own tests, the agent used 53% less code, 41% less time, 26% lower cost, and 45% fewer tokens. It says 98% of risky logic shipped with a test, against 68% without the skill. The test used 39 tasks, five runs each. These are project-reported results and we could not verify them. The Humanizer README likewise reports that judges preferred its rewrites 16 times out of 16, and it also says its output is not meant to fool AI detectors and often still gets flagged.

Anthropic’s own README adds a warning. It says its skills are provided for demonstration and education and should be tested before you rely on them.

Why stars are not trust

A skill is not just text. It can tell an agent to run scripts, read files, and call websites. Anthropic’s documentation says to use skills only from trusted sources, meaning ones you wrote or got from Anthropic. It warns that a malicious skill can steer the agent into running code that does not match what the skill claims to do. Anthropic’s help pages add that you should review a skill from any less-trusted source before enabling it.

By that advice, nine of these ten need a review before you install them. A star count cannot tell you whether anyone has done that review. Here is a short checklist:

  1. Open SKILL.md and read it. Look for lines that send data to a web address.
  2. Open any script in the folder. Check what files it touches.
  3. Pin the version you reviewed. A repository can change after you read it.
  4. Try it first in a throwaway project.

What we could not check

  • The star counts are one snapshot and change constantly.
  • Star counts show interest, not use, quality, or safety.
  • We did not install or test any skill, so we make no claim that they work as described.
  • Our topic search could miss a popular skill that has no matching tag.

Sources