Skip to content
Agent Skills Guide
Try Kortix
Contents

The Agent Skills format, from the file up

How AI agent skills earn a load in an open-source repo

A skill works when the agent can find it in the repo, read a description that matches the task, and load it under a rule a person approved. Kortix is open source, and it runs that path end to end, from the description line to the grant that allows a load.

Try KortixRead the specification
The Agent Browser skill's page in the Kortix marketplace, with the description an agent reads before it loads the skill
https://kortix.com/marketplace/kortix-starter/agent-browser
---name: kortix-branddescription: "Load FIRST for anything that carries the Kortix look or voice: product or mobile UI, copy of any kind, decks, social, images, email, CLI output, anything with the logo, and reviews of these. Routes you to the exact files. Triggers: choosing a spacing, color, radius, shadow, duration or word that a user or buyer will read."---

https://github.com/kortix-ai/suna/blob/dev/.agents/skills/kortix-brand/SKILL.md

Open source. Read it, fork it, audit it.

Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. · Session work reaches main through a change request · Merge is default-deny for agents

Three skills, three owners

One a team writes for itself, one a company ships to every agent, one an agent installs and pins.

Team rule

Skill file

The team's review checklist, kept where the agent can find it.

The team that owns a release keeps its four checks in a skill, so an agent that touches a release reads the same steps a person does.

skills/release-check/SKILL.md

  • Migration addedyes
  • Rollback stepyes
  • Changelog entrymissing
  • Feature flagyes

3 of 4 checks pass

Brand voice

Skill file

The house voice, shipped to every agent that writes.

Marketing keeps one brand skill in the repo, so support replies, release notes and email land in the same voice without a meeting.

skills/brand-voice/SKILL.md

ChannelRuleExample
EmailPlain sentencesYour invoice is attached
Release notesName the change firstExport now retries twice
Error copySay what failedCould not save the draft

One voice on three surfaces

Marketplace

Lockfile

A skill an agent installs, then pins by hash.

An agent pulls a skills folder from a public repo and records the source and hash, so the next session runs the same bytes.

skills-lock.json

"skills": { "pdf-forms": {"source": "github.com/acme/skills", "skillPath": "pdf-forms", "computedHash": "9f2c41…"},}

One skill added, pinned by hash

Where a skill lives decides whether it loads

The format is the same in all three places. Discovery and permission are where they differ.

Kortix

Open source
Yes
Where a skill lives
skills/<name>/SKILL.md in your project repo
Who can load it
A skill in skills/ loads only when that agent's grant allows it
How a change is reviewed
A change request a person approves
What loads into context
Metadata at startup; Kortix does not put skill bodies in every prompt
Who owns the format
The Agentic AI Foundation, under the Linux Foundation

A folder in your repo

Open source
Yes
Where a skill lives
A directory the agent's filesystem can read
Who can load it
Any agent that can read the directory
How a change is reviewed
Whatever review your git host enforces
What loads into context
Whatever the host agent's loader chooses to read
Who owns the format
The Agentic AI Foundation, under the Linux Foundation

A vendor skill library

Open source
Yes for the format, the library is hosted
Where a skill lives
The vendor workspace, or ~/.claude/skills/
Who can load it
Every member of the API workspace; claude.ai copies stay personal
How a change is reviewed
A new version uploaded to the workspace
What loads into context
Metadata at startup, body under 5,000 tokens when triggered
Who owns the format
The Agentic AI Foundation, under the Linux Foundation

One format, one foundation, and a repo you own

Anthropic handed the format to a neutral foundation, and more than 70 platforms ship it.

  • Donated: Anthropic gave Agent Skills, its name, its logo, its GitHub organization and the agentskills.io domain to the Agentic AI Foundation, which it co-founded with Block and OpenAI; the Model Context Protocol went there too. announcement
  • Maintained: engineers from Anthropic, Google, OpenAI and other companies maintain the standard as individuals, with David Soria Parra and Mahesh Murag as lead maintainers.
  • Shipped: more than 70 platforms have integrated Agent Skills, including ChatGPT, Cursor, GitHub Copilot and VS Code.
  • Ours: Agents, skills, memory, tools, triggers and connectors are files in one git repo you own. Read the docs

Questions people ask about agent skills

What are examples of skills in AI agents?
A release checklist, a brand voice guide and a PDF form filler are three. Each is a folder with a SKILL.md: YAML frontmatter that names it and says when to use it, then Markdown steps an agent follows.
Are agent skills actually useful?
Yes, when the agent can find and load them. A skill keeps a procedure out of every prompt: metadata loads at startup and the body loads only when a task matches, so a large library costs little context.
What skills do you need to build AI agents?
Start with three: one for how your team reviews output, one for how you write, and one for the tool your agents touch most. Write the description line first, because that is what the agent matches a task against.
Is agent skills replacing MCP?
No. MCP connects an agent to outside systems, and a skill teaches it how your company does one job. They run together: an agent can load a billing skill and reach the billing API through an MCP server.

Give your agents skills they can find.

Agents, skills, memory, tools, triggers and connectors are files in one git repo you own.

Try KortixRead the docs

Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. · Session work reaches main through a change request · Merge is default-deny for agents