Agent skills study 2026: what 10,000 popular skills do
10,000 agent skills from 2,841 public GitHub repositories with 100 to 10,000 stars, sorted by the job each does and the field it serves, with the AI products they name, how they are built and their licences.
Data collected 20 November 2026. Next collection November 2027.
0:00
13:42
skills examined
10,000
repositories with 100 to 10,000 stars
2,841
made for a field outside technology
34%
of repositories with no licence detected
12%
Main findings
Writing and changing code is the most common job, at 1,314 (13.1%) of the 10,000 skills, followed by building and running AI agents at 10.9%.
33.5% of the skills are made for a field outside technology; Science and research leads with 828 (8.3%).
Popularity differs little by job: the median repository holding each kind of skill has between 280 and 397 stars. By field, health and sales skills sit in the most-starred repositories (437 and 435).
Claude is named in 21.8% of the skills, Codex in 7.3% and Gemini in 3.2%; 73.2% name no AI product at all.
353 (12.4%) of the 2,841 repositories carry no licence GitHub can detect, so 14.3% of the skills come with no stated permission to reuse them.
Every figure describes the 10,000 skills examined on 20 November 2026, from repositories with 100 to 10,000 stars. Repositories with more than 10,000 stars are reported apart.
What the skills do
Base: 10,000 skills from 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026. One main job for each skill.
Main job, with the repositories that hold such skills and their median stars (base: 10,000 skills)
Main job
Skills
Share
Repositories
Median stars
Write and change code
1,314
13.1%
610
349
Build and run AI agents
1,087
10.9%
637
308
Work with data
880
8.8%
372
287
Plan and manage work
777
7.8%
469
291
Research and gather information
759
7.6%
368
343
Deploy and run systems
685
6.8%
431
356
Test and fix bugs
656
6.6%
436
334
Check security
638
6.4%
229
309
Run business tasks
569
5.7%
219
328
Write and edit text
474
4.7%
326
342
Market and sell
429
4.3%
189
348
Design interfaces
428
4.3%
281
307
Review code
422
4.2%
317
345
Make images, video and audio
352
3.5%
239
328
Make documents and slides
220
2.2%
158
397
Other
156
1.6%
93
284
Teach and explain
154
1.5%
100
280
Which fields they serve
Base: 10,000 skills from 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026. One field for each skill. "No particular field" covers skills of use to anyone.Base: 10,000 skills from 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026. Fields outside technology only: 3,350 (33.5%) skills.
The most common jobs within fields outside technology (base: 10,000 skills)
Field
Main job
Skills
Science and research
Work with data
337
Marketing
Market and sell
313
Science and research
Research and gather information
226
Design and creative work
Make images, video and audio
220
Finance
Run business tasks
141
Legal
Run business tasks
115
Finance
Work with data
96
Media and publishing
Write and edit text
95
Design and creative work
Design interfaces
95
Science and research
Write and edit text
75
Popularity by job and field
Median stars of the repositories that hold at least one skill with each main job. Jobs found in fewer than 20 repositories are left out. Base: 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026.Median stars of the repositories that hold at least one skill for each field. Fields found in fewer than 20 repositories are left out. Base: 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026.
AI products named
Base: 10,000 skills from 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026. A skill can name several products.
1,392 (13.9%) skills name Claude and no other AI product. The Model Context Protocol (MCP) is named in 1,544 (15.4%).
How the skills are built
Base: 10,000 skills from 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026.Base: 8,284 skills kept in a folder of their own, collected 20 November 2026. A skill can appear in more than one row.
The median skill runs to 133 lines, and the median description to 229 characters. 169 (1.7%) skills have exactly the same text as a skill in another repository.
Licences
Base: 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026. Licence as detected by GitHub for the whole repository.
2,025 (71.3%) repositories carry a permissive licence such as MIT or Apache 2.0. A skill in a repository with no licence comes with no stated permission to copy or change it.
Stars and activity
Base: 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026.Time from the last commit on the main branch to 20 November 2026. Base: 2,841 public GitHub repositories with 100 to 10,000 stars, collected 20 November 2026.
1,992 (70.1%) of the repositories were created in 2026, and 25 (0.9%) are archived.
Repositories over 10,000 stars
424 repositories with more than 10,000 stars hold 26,648 SKILL.md files. They are kept out of the 10,000 so that a few very large collections do not decide the figures. A separate sample of 500 skills from 127 of them was sorted the same way: build and run ai agents is the main job of 74 (14.8%).
Main job in repositories over 10,000 stars (base: 500 skills)
Main job
Skills
Share
Build and run AI agents
74
14.8%
Check security
55
11.0%
Write and change code
50
10.0%
Market and sell
32
6.4%
Deploy and run systems
32
6.4%
Work with data
32
6.4%
Method
Repositories. GitHub repository search by topic and keyword (agent skills, Claude skills, Claude Code, Codex, Gemini CLI and others), and GitHub code search for files named SKILL.md, gave 44,245 candidate repositories. Forks were left out. 10,182 repositories with 100 to 10,000 stars held at least one SKILL.md file, 359,868 in all.
Skills. Every file named SKILL.md in each repository's full file list. To keep a few large collections from deciding the figures, each repository gave at most 40 skills, chosen in a fixed random order. 10,000 skills were then taken in the same order. Unreadable files and files that only point to another SKILL.md were replaced by the next in order.
Job and field. Each skill was sorted into one of 17 jobs and one of 22 fields by a language model (gemini-3-8-flash) from its name, description, opening text and repository description, using fixed lists. The same model marked empty templates and stubs.
Header. Read from the block between two lines of three hyphens at the top of the file. "Meets every header rule" means a name in lower-case letters, digits and hyphens of up to 64 characters, equal to the folder name, and a description of 1 to 1,024 characters, as the Agent Skills specification sets out.
Scripts. A skill ships a script when its folder, or a folder directly inside it, holds a file ending in .py, .sh, .bash, .zsh, .ps1, .bat, .cmd, .js, .mjs, .cjs, .ts, .rb, .pl, .php, .lua, .r. Skills at the top of a repository are left out of this count.
AI products. A search of the full text for each product name, without regard to capitals.
Licence, stars and activity. As GitHub reports them for the repository on the day of collection.
Check of a sample
400 skills, every 25th in a fixed random order, were read in full by a second reader (a second AI model, Claude Sonnet, not a person) and compared with the recorded values.
Disagreements found (base: 400 skills)
Value
Wrong
Error rate
Main job
0
0.0%
Field
2
0.5%
Template or stub
2
0.5%
Name and description
0
0.0%
Ships scripts
0
0.0%
AI products named
5
1.3%
In all, 9 (0.4%) of 2,400 values checked were wrong. Figures are given as counted, without correction.
Limits
GitHub search finds repositories by their topics, names, descriptions and recent activity. Repositories that hold skills but are described in other terms may be missing.
Stars belong to a repository, not to a skill. Installs and downloads are not public, so use of a single skill was not measured.
The job and field come from a language model; the check above gives their error rate.
No skill was run, and no script was read or tested for safety.
The figures describe one day. Repositories change, and the next collection will differ.
Data and citation
One record for each of the 10,000 skills, with no repository, owner, skill name or file path: repositories are numbered, stars are given in bands and the last commit by month.
Published by Knacco, a library of ready-to-use AI skills, prompt packs and workflows. The data, figures and charts are free to share and adapt under the CC BY 4.0 licence, with credit and a link to this page. Corrections are received through the contact page.
Agent skills study 2026: what 10,000 popular skills do. Data collected 20 November 2026. https://knacco.com/en/research/agent-skills-study