Skip to content
Agent Skills OS

Docs

Write a skill once. Load it into any agent.

A skill is one Markdown file. The package reads it and hands each runtime exactly what it expects: a system prompt, an assistant config, a prompt template, a rule file.

npm install universal-agent-skills

01 — Adapters

One file in, the right shape out.

Pick a target to see what its adapter returns for this example skill. Every adapter reads the same file, so there's nothing to keep in sync.

skills/release-notes/SKILL.mdExample
---
name: release-notes
description: Turn merged pull requests into release notes grouped by what users notice.
version: 0.1.0
tags: [writing, engineering]
inputs:
  - name: changes
    description: Merged PR titles and descriptions.
    required: true
---

# release-notes

Group changes under New, Improved and Fixed.
Write one line per change, in plain words, present tense.
Leave out internal refactors unless they change behavior.
Returned value

forClaude("release-notes")Pass system to anthropic.messages.create()

{
  "system": "# release-notes\n\nGroup changes under New, Improved and Fixed.\nWrite one line per change, in plain words, present tense.\nLeave out internal refactors unless they change behavior.",
  "metadata": {
    "skill_name": "release-notes",
    "skill_version": "0.1.0",
    "related_skills": []
  }
}

02 — Usage

Drop it into the code you already have.

Each adapter is one function call. Python frameworks read the same Markdown files straight from node_modules.

agent.js
import Anthropic from "@anthropic-ai/sdk";
import { forClaude } from "universal-agent-skills";

const anthropic = new Anthropic();
const { system } = forClaude("super-ui");

const msg = await anthropic.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 4096,
  system,
  messages: [{ role: "user", content: "Review this dashboard screen." }],
});

03 — API

Fourteen functions. That's the whole API.

Everything takes a skill name and returns plain data. Types ship in src/index.d.ts.

Load

loadAllSkills()
Skill[]
getSkill(name)
Skill
listSkills()
SkillSummary[]
getRelatedSkills(name)
RelatedSkillSummary[]
asSystemPrompt(name)
string

Adapt

forClaude(name)
{ system, metadata }
forOpenAI(name)
{ name, description, instructions }
forLangChain(name)
{ template, inputVariables }
forCrewAI(name)
{ role, goal, backstory }
forAutoGen(name)
{ name, system_message }
forCursor(name)
{ filename, content }
forAgentsMd(name)
{ filename, content }
forStitchPrompt(name)
{ content }
forGeneric(name)
{ name, description, version, … }

04 — Write a skill

A folder and a Markdown file.

Put a SKILL.md in skills/<name>/. There's no registry and no build step. Add a references/ folder for longer material the skill points to.

skills/my-skill/SKILL.md
---
name: my-skill
description: When to use this skill and what it produces.
version: 0.1.0
tags: [domain, capability]
inputs:
  - name: input_name
    description: What this input is.
    required: true
related: [other-skill]   # optional
author: your-name        # optional, only if you wrote it
---

# my-skill

The instructions the agent follows: the process,
the output shape, and what to avoid.
  1. Lead with intent. The first paragraph says when the skill applies.
  2. Process before output. Describe how to think, not only what to emit.
  3. Specify the output shape. Models drift without a target format.
  4. Name the failure modes. A "what to avoid" section prevents the common mistakes.
  5. Be concrete. One real example teaches more than a paragraph of advice.
Contribute it to the library