How we assess skills
The library collects skills from public GitHub repositories, evaluates each SKILL.md against the same criteria and shows only those that pass every threshold.
Skill sources
Skills were collected in three rounds. Imported content comes from its author’s GitHub repository, together with the licence file. Other catalogues supplied repository addresses only, from pages allowed by robots.txt.
Round 1: 25–26.10.2026
The starting point was skillsclau.de (33,971 skills from 3,957 repositories), followed by complete skill folders downloaded from those repositories: 333,885 SKILL.md files. Removing identical copies, artefacts (tests, backups, archives) and near duplicates left 236,717 canonical skills.
| Stage | Skills |
|---|---|
| SKILL.md files in repositories | 333 885 |
| Canonical skills (without copies or artefacts) | 236 717 |
| “Good” class | 89 217 |
| Good, current and without detected risk signals | 82 798 |
| Plus an open licence | 64 025 |
| Without attack or penetration testing skills | 61 415 |
| With a complete downloadable package | 61 292 |
| After rescanning all scripts (27.10.2026) | 61 277 |
Round 2: 27–28.10.2026
The second round covered repositories listed by other catalogues and collections: skills.sh, claudemarketplaces.com, SkillsMP, Smithery, agentskill.sh, claudepluginhub.com, awesomeskill.ai, geminiskill.org, GitHub “awesome” lists, Claude plugin marketplaces and official company repositories (Anthropic, OpenAI, Google, Oracle, Vercel, Microsoft, Stripe, Cloudflare and others). Repositories from the first round were skipped. We checked 20,391 repositories; 17,773 contained skills.
We excluded bulk collections: repositories that copy other people’s skills wholesale or generate them at scale (registries, mirrors, “awesome” collections and archives of removed skills). Originals from these collections are usually already included from their authors’ repositories. Official repositories are exempt from this rule.
| Stage | Skills |
|---|---|
| SKILL.md files in repositories | 2 004 982 |
| Without folders identical to those already downloaded | 1 234 732 |
| Without bulk collections or repositories lacking an open licence | 183 029 |
| Without artefacts | 180 381 |
| “Good” class | 98 889 |
| Good, current and without detected risk signals | 87 278 |
| Plus an open skill licence | 85 816 |
| No near duplicates of skills already in the library or this round | 80 904 |
| Without attack or penetration testing skills | 80 091 |
| With Polish summaries, added to the library | 79,276 |
Round 3: 13–14.11.2026
Round three covered GitHub repositories created since 26.10.2026 with Agent Skills topics and descriptions, new addresses from catalogues, official company repositories, the Polish shelf, and library repositories that changed since the previous round. We checked 34,693 repositories; 28,369 contained skills.
When the same content appears in several repositories, the author is the repository where SKILL.md appeared first (date of the first commit). A bulk collection never wins over the author's repository. Under this rule, 1,572 library skills now point to the author's repository and keep their existing library address.
| Stage | Skills |
|---|---|
| SKILL.md files in repositories | 1 345 343 |
| Without folders identical to those already downloaded | 223 747 |
| Without bulk collections | 150 705 |
| “Good” class, current, no risk signals, open licence (including author versions from rounds 1 and 2) | 79 204 |
| Without near-duplicates, author first | 55 846 |
| Without attack or penetration testing skills | 53 922 |
| With Polish summaries and after scanning all scripts, added to the library | 52,713 |
Total in the library: 193,267 skills. New skills from round three have a catalogue addition date of 13.11.2026.
First-party skills
The library also includes skills published directly by Bartosz Sokoliński. They have source packages, open licences, the same format assessment and script scanning. Their source is this website.
Quality score (0 to 100)
A script calculates the score from SKILL.md and its repository. Criteria follow Anthropic’s Agent Skills best practices (platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices) and the format specification (agentskills.io/specification).
| Component | Points | What we check |
|---|---|---|
| Format | 25 | Frontmatter with name and description (10), valid name (5), description up to 1,024 characters (5), name matching the folder (5). |
| Discoverability | 25 | Description explains when to use the skill (12), has 80–1,024 characters (8), and uses third person (5). |
| Content | 35 | At least 150 words and at most 500 lines (12), at least two headings (8), step lists (5), examples or code (10). |
| Hygiene | 15 | No placeholders such as TODO (8), no Windows paths (2), with a licence (5). |
The “good” class requires a score of at least 80, a description explaining when to use the skill, a licence, at least 150 words and no placeholders. Outdated skills are excluded: those marked deprecated, referencing retired models or APIs, or from repositories unchanged for over six months.
The score measures format and completeness, rather than how well a skill solves a problem. Random samples from each class were also reviewed manually. “Bartosz’s Pick” is a shortlist chosen by the library author.
Risk scan
Skills may contain model instructions and scripts that Claude can execute. The scan searches SKILL.md and scripts for abuse signals. Any detected signal excludes a skill from the library.
Since 27 October 2026 (version 1.3.0), the scan covers every script in a published package, including subdirectories and .ts, .tsx and .jsx files. The original assessment checked up to 30 scripts per skill and, in some repositories, only files alongside SKILL.md. Rescanning removed 15 skills and corrected the script count for 351. A skill with a script larger than 2 MB that cannot be checked is excluded. The scanner and its rules are in the website repository (skille-narzedzia/skanuj_paczki.py).
- Executing decoded code (base64 to shell, eval or exec of decoded text), reverse shells.
- Reading SSH keys, system passwords or cloud credentials; accessing the macOS keychain.
- Phrases overriding model instructions outside quotations; hidden bidirectional text control characters.
- Sending data to capture services (webhook.site, requestbin, pipedream, interact.sh, Discord webhooks) or ngrok tunnels.
- Downloading and running scripts in a single command (curl to shell), or deleting the home directory.
- HTML comments containing instructions for the model.
The library excludes skills for carrying out attacks: penetration testing, red teaming, exploits, bypassing safeguards, jailbreaks and OSINT about individuals. Repositories primarily dedicated to these topics were also excluded. Defensive skills, configuration reviews and secure coding remain in Security when they do not cover attack techniques.
The scan uses patterns and cannot detect everything. Review SKILL.md and especially scripts before installing. The library has not executed package contents. Executables (such as .exe, .dylib, .so and .jar) are excluded from .zip packages.
Licences
The library includes only skills from repositories with an open licence recognised in a LICENSE file. Repositories without a licence (all rights reserved), with non-commercial licences (CC-BY-NC), custom terms or unrecognised licences were excluded. Each .zip includes the repository licence and ZRODLO.txt with the source address.
| Licence | Skills | Meaning |
|---|---|---|
| MIT | 148,240 | Use and modify while retaining author attribution and licence information. |
| Apache-2.0 | 33,240 | Like MIT, plus change notices and a patent licence. |
| Other open licences | 6,763 | Pozostałe otwarte: BSD, GPL, LGPL, MPL, CC-BY, CC0, Unlicense. |
| AGPL-3.0 | 5,024 | Copyleft: modifications distributed to others, including over a network, must use the same licence. |
This is information, not legal advice. The licence text in the author’s repository governs use.
Connecting through MCP
The MCP server lets Claude search the library in conversation, read SKILL.md and install a skill with your consent. It is read-only and requires no login.
Server URL: https://sokolinski.com/skille/en/mcp
Claude.ai and the Claude app
- In settings, open Connectors and add a custom connector.
- Enter a name, such as “Skills Library”, and the server URL above. Leave authentication blank.
- Enable the connector in your conversation and describe your task, for example: “find a skill to review an AI supplier contract”.
Claude Code
claude mcp add --transport http biblioteka-skilli https://sokolinski.com/skille/en/mcp
Server tools
| Tool | What it does |
|---|---|
| szukaj_skilli | Search by task description in Polish or English, with filters for area, task type, licence, no scripts, Polish collection and minimum score. |
| pokaz_skill | Full card: descriptions, author, licence, score breakdown, scripts, files and SKILL.md content. |
| pobierz_skill | Package files and contents, .zip URL, SHA-256 and installation path. Call only after user consent. |
| sprawdz_aktualizacje | Compare installed skill versions with the database. |
Polish summaries and categories
A language model (Claude) generated the Polish summaries, work areas and task types from authors’ original descriptions. Each skill page shows the original. If a Polish summary differs, the original takes precedence.
For whom: organisation types
Since version 1.4.0, each skill has one to four tags for the kinds of organisations where it may be useful. A language model (Claude) assigned these using the skill description, name and work area. These tags help you browse. They do not establish compliance with industry requirements, such as public procurement or medical documentation rules.
| Organisation type | Example skills |
|---|---|
| Startup | MVPs, pitch decks, fundraising, growth and rapid prototypes. |
| Software companies & IT | Programming, testing, DevOps, architecture and cloud. |
| Small & medium businesses | Invoices, online stores, customer service, CRM and simple office automation. |
| Large companies & corporations | Processes, management reporting, governance, HR and enterprise systems. |
| Banking, insurance & finance | Risk, accounting, financial analysis, regulation, payments and KYC. |
| Public administration & local government | Official correspondence, public procurement, plain language and digital accessibility. |
| Universities & research institutes | Research, literature reviews, statistics, publications and grants. |
| School | Lesson plans, student materials, quizzes, assessment and language learning. |
| Healthcare | Medical documentation, clinical guidelines, pharmacy and research. |
| Law firms & legal departments | Contracts, legal reviews, GDPR, regulation and compliance. |
| Non-profit organisations | Grant applications, fundraising, volunteering and reporting. |
| Media, agencies & creators | Content, video, graphics, SEO, social media and campaigns. |
| Industry, trade & logistics | Manufacturing, supply chains, warehousing, e-commerce and quality. |
| Freelancers & sole traders | Productivity, proposals, billing and solo work. |
Reports and removals
If a skill contains harmful content, differs from its description or violates a licence, report it through the feedback form or email bartosz@sokolinski.com and include the skill page URL.
Authors who do not want their skill in the library can request removal through the form or from that email address. The skill will be removed at the next database update, or immediately for harmful content.
Ranking and statistics
The website anonymously counts skill views and downloads (.zip, SKILL.md and MCP). Each address counts once per skill per day, identified indexing bots are excluded and the counter database stores no IP addresses. Ranking becomes public after 100 downloads within 30 days. Details in privacy policy.