Skill Seekers — builds your own skill from external documentation
You have a tool you know well, and you want your assistant to know it just as well. Skill Seekers turns that tool’s documentation into an instruction file the agent reads before it answers.

What it does
You have a tool you know well. But you want your assistant to know it just as well. Skill Seekers takes that tool’s documentation and turns it into an instruction file the assistant reads before it answers you.
Think of it as a photocopier for knowledge. You give it a documentation site, a repository, a PDF, or a video. It gathers the pages, sorts them by topic, and writes one unified text file, SKILL.md.
SKILL.md is not a random note. At the top it has two fields, name and description, in a format the assistant recognizes without any extra setup.
The sources it accepts are eighteen. Documentation sites, code repositories, PDF, Word, EPUB, Jupyter notebooks, OpenAPI, PowerPoint, RSS, man pages, videos from YouTube or Vimeo, and pages from Confluence, Notion, and Slack. Whatever it collects, it rewrites into twenty-two output formats (repository).
It is a command-line program. Command line means you work with it by typing commands into a black window, not by clicking buttons. It also has an optional MCP server, which is a way for the assistant to call it on its own from inside itself.
What it does not do
Do not make it run or verify code — we tested the heart of the flow and saw what it does well. But it is not a tool that executes or verifies code. Do not ask it to run examples, do integration tests, or confirm that a snippet works in production. SKILL.md is a summary, not an executable artifact. For small text fixes it is not worth the effort — you will waste time on something that only does copy-rewrite.
Do not expect automatic, one-click installation in every assistant. The tool has ready-made flows for Claude and Cursor, but for other assistants the packaging and installation require manual steps. We did not fully test the AI enhancement flow, and using it with third-party engines may require API keys and extra setup.
Do not underestimate the sources: it does not scrape content behind authentication or heavy JavaScript very well without extra setup. The trap is to treat SKILL.md as an absolute truth. It is a snapshot and can compress or omit details that still need human review, especially in sensitive API or security sections.
Who it is for
For anyone who has a tool, a library, or a project and wants their assistant to speak with knowledge, not guesses.
It is useful when the documentation runs to many pages. It is useful when you work with code and want the assistant to know the library’s patterns.
It is not for a small text correction. There it is not worth the effort.
Installation
You need Python 3.10 or newer, and Git. Otherwise, one command.
uv tool install skill-seekers
Check that it was installed correctly:
skill-seekers --version
You should see a version number. In our test it returned “skill-seekers 3.10.0”.
If you also want the MCP server, install the same tool with the extra:
uv tool install "skill-seekers[mcp]"
Installation with pip
If you prefer pip, the same applies. It also needs Python 3.10 or newer.
pip install skill-seekers
Use in each LLM
The heart of the flow has two steps: build the skill, then package it for your assistant. The second step changes depending on the program you use. This is the part that is usually missing from articles.
Claude Code
Building starts with one command. Run it from the folder where you want the result to appear.
skill-seekers create https://docs.djangoproject.com/ --target claude
After packaging:
skill-seekers package output/django --target claude
You get a zip file. That is the skill ready for Claude Code.
You have two ways to get it there. The first is to put it in Claude Code’s skills folder by copying the skill directory:
cp -r output/django ~/.claude/skills/
The second is to upload it from Claude’s page at claude.ai/skills.
If you want Claude Code itself to drive packaging, set it up as an MCP server in stdio mode:
python -m skill_seekers.mcp.server_fastmcp
And one last thing: you can let Claude Code itself write the skill enhancement, without a separate API key. The documentation gives this with –enhance-mode local.
Codex CLI
Here I need to be careful, and I will be honest with you. Skill Seekers uses Codex CLI as an engine, not as a final destination. That means you tell it to run enhancement with Codex, and it works on your subscription instead of billable tokens:
skill-seekers create <source> --agent codex
In the list of assistants that accept installation, Codex does not appear. The documentation gives a separate folder for Claude Code, Cursor, and others, but not for Codex. So here I am talking about using Codex as the engine, not about installation inside it. That the generated SKILL.md is also readable by Codex is my own estimate, not confirmed by the sources.
Cursor
Cursor has its own path, and it is the cleanest of the three.
To drop the skill into the project, one command:
skill-seekers install-agent output/react/ --agent cursor
The documentation says this writes into the project’s .cursor/skills/ folder.
If you want Cursor to call Skill Seekers on its own, run it in HTTP mode:
python -m skill_seekers.mcp.server_fastmcp --transport http --port 8765
And if you prefer the old .cursorrules file, export the content as plain markdown and copy it there. The documentation recommends that too.
Where it wins, where it loses
It wins on breadth. Eighteen sources and twenty-two outputs mean you do not gather the same material again for each assistant. You gather it once and export it wherever you want.
It also wins on price. It is MIT, free, and runs locally. The basic flow with –enhance-level 0 works without an API key.
It also wins on depth. It reads a repository’s code, extracts design patterns, and writes usage guides.
It loses on terminology. It is built around the Claude ecosystem, and that shows in the names, defaults, and functions. Anyone working with another assistant will feel they are moving through someone else’s house.
It also loses on time. A small documentation set takes five to ten minutes. A large one can take up to an hour (repository).
And it loses a bit on the installation experience inside an assistant itself. As you saw with Codex, it does not have its own destination for every program.
Test
It really ran, on this machine, with uv.
The installation went through cleanly, and the version command returned “skill-seekers 3.10.0”.
Then a real test ran against a documentation site:
skill-seekers create https://docs.astral.sh/uv/ --name uv-probe --max-pages 5 --enhance-level 0
It passed. A folder output/uv-probe/ was created with SKILL.md 2.127 bytes, with the correct name and description fields at the top. That is proof that the heart of the flow works without an API key.
What was left out. We did not run the packaging step. We did not run the AI enhancement flow. We did not run the server mode as an MCP server. What I write above about those three comes from the documentation, not from hands-on testing.
Verdict
Skill Seekers solves a problem we all have and few name: how to give your assistant solid knowledge about the tool you work with. It does it for free, locally, and with a broad source list.
It is mature for the basic flow. The build works, and we saw it. It is weaker when you ask it to work with something beyond Claude, and that is where you expect the project’s next step to be.
If you work with Claude Code or Cursor and have a tool you want it to know well, try it. It will save you hours.
Score: 8.6/10.
Argos’s score for the tool choice: 86/100.
See also
In the same category we also tested:
- openwiki: the repository wiki that writes and maintains itself
- graphify: the knowledge graph that answers for your repository without rereading it
- Docling: your documents in clean Markdown, with the structure intact
Sources:
repository https://github.com/yusufkaraaslan/Skill_Seekers
documentation https://skillseekersweb.com/docs/getting-started/quick-start
releases https://github.com/yusufkaraaslan/Skill_Seekers/releases
package page https://pypi.org/project/skill-seekers/