“I know kung-fu.”
A skill is a SKILL.md with instructions, scripts and resources an agent loads only when the task needs it: a PDF, a spreadsheet, a presentation, your deployment, your checklist. We define skills for your processes, curate the best ones in a library and improve them from every run.
Agent skills
Instructions, scripts and resources in a SKILL.md your agent loads when it needs them.
What we do
Defined from your process
We watch how the job is done by hand, then write the steps, checks and limits for an agent. Short, operational, testable.
A curated library
The skills we run in production, versioned with a changelog. You can read ours; yours stay private.
Self-improving
After a run, the agent leaves a note about what surprised it. The notes become a revision, a new version and a changelog line, reviewed by a person.
Portable
The open Agent Skills format works in Claude Code, Codex, Hermes, Pi and any agent that reads a SKILL.md.
How it works
- 1
Pick the process
Something done repeatedly by people or agents, with mistakes that cost time.
- 2
Write and test the skill
Instructions, scripts and resources, tested against real runs until the agent gets it right without help.
- 3
Ship to your agents
Installed where your agents live, loaded only when the task needs it.
- 4
Evolve
Run notes in, revised skill out. A monthly pass keeps the library sharp.
What you get
- A skills repository for your processes
- A library page with versions and history
- The improvement loop set up and documented
- A session to teach your team to write the next ones
Good fit if
- Processes done repeatedly, by people or by agents
- Teams already using Claude Code, Codex, Cursor or Hermes
- Knowledge that lives in people's heads and should not
Not a fit if
- One-off tasks
- Processes nobody can describe yet
Stack we use
Questions
Can we see examples?
Yes. Our public library lists the skills we use on our own site and client work, with their SKILL.md and changelog.
Do skills work with any model?
The format is model-agnostic. We test ours with Claude Code and Pi, and agents on other models read the same file.
How do they improve?
Agents write a run note when something was wrong or missing. A review step proposes a revised SKILL.md and a changelog line; a person merges it and the version goes up.