Agent Skills left Anthropic. Write yours like this

On 22 October Anthropic donated Agent Skills to the Agentic AI Foundation, under the Linux Foundation. The name, the logo, the GitHub org and agentskills.io go with it. More than 70 products already load the format, including ChatGPT, Cursor, GitHub Copilot and VS Code. Anthropic says a million skills have been published.

That number should make you careful, not excited.

We run a library of more than 40 skills on our own work: product backends, video pipelines, reviews, deploys. A skill here is a written playbook an agent must follow, not a bag of tips. The donation does not change that. It does mean the format is no longer one vendor’s side project.

Key takeaways

• A skill is a folder: SKILL.md plus the scripts and notes for one job.
• The standard now lives at the AAIF, next to MCP. You can keep using it without joining.
• Short, scoped skills beat long ones. If it cannot name the files it touches, it is not ready.
• Treat a downloaded skill like untrusted code. Read it before any agent does.
• Write the skill after you have done the job twice by hand. Not before.

What a skill actually is

The spec is small on purpose. One folder. A SKILL.md that says when to load it and how to do the job. Optional scripts and reference files sitting next to it.

That is the whole trick. The agent does not need your entire company in context. It needs the procedure for this task, loaded when the task starts.

We built ours the same way Anthropic describes: general agents are fine at “write a function.” They are weak at “this repo’s review loop, this product’s brand rules, this deploy command and no other.” Those rules used to live in a person’s head, or in a 4,000-line prompt that nobody updated. A skill is the version you can read, diff and revert.

Why the donation matters, and why it does not

MCP moving to the same foundation last year was the template. A standard that only works inside one lab dies the first time you switch tools. Skills already jumped: loaders showed up for Codex, Cursor and OpenCode within weeks of the first public launch.

A Linux Foundation home makes it harder for one company to fork the format into a product lock. Engineers from Anthropic, Google and OpenAI will maintain it as individuals. That is the right shape.

It does not make a random SKILL.md on GitHub safe. It does not mean you should install 50 of them. Anthropic’s own next-step list includes versioned releases, a proposal process and security guidance for people who write and install skills. That last item is the one most teams will feel.

How we write a skill so it stays useful

The bad ones read like a wiki dump. The agent skims, misses the constraint, and does the job the way it always does.

The ones that last look like this:

1. One job. “Upload a YouTube Short to this channel” is a skill. “Do all marketing” is not.
2. A trigger line at the top. When to load it, in one sentence. If two skills could fire on the same request, rewrite them until they cannot.
3. Exact paths and commands. Named files, named tests, named deploys. No “the usual script.”
4. Hard stops. Brand rules, secrets, “never push,” “never mention AI on this channel.” Put them near the top. Agents obey the first screen more than the last.
5. A way to check it worked. A test, a URL to open, a log line that must appear. If you cannot say how you would know, you cannot hand it to an agent.

We only write the skill after the job has been done twice. The first time you invent the steps. The second time you notice what you skipped. Then you write it down. That matches how we ship software with agents: spec, small tasks, someone else reviewing.

Keep the file short enough to load. Long reference material goes in sibling files the skill can open when needed. The agent should not pay for the whole manual on every turn.

What not to do this week

Do not scrape a thousand public skills into your repo. You would not curl a stranger’s bash into production. A skill with a scripts/ folder is that curl, with extra English on top.

Do not put API keys, cookies or private customer text in the skill folder. The agent will eventually print them.

Do not write a skill that marks its own work as done. Review stays a separate step, preferably a separate skill, run by a different agent or a person. That is the same rule as independent evals.

Do not wait for the AAIF roadmap before writing the ten skills your team already repeats. The format is stable enough. The security guidance will catch up. Your procedures will not write themselves.

The bottom line

Agent Skills leaving Anthropic is good news if you already treat them as versioned playbooks. It is noise if you treat them as a plugin store.

Write the job. Name the files. Put the hard stops first. Load it only when that job starts. That is the standard, foundation or not.

We build scoped agent workflows and written skills as AI agent development. If you want that set up in your repo, write to us.

FAQ

What did Anthropic donate on 22 October 2026?

The Agent Skills standard: the name, logo, GitHub organisation and agentskills.io domain. It now sits at the Agentic AI Foundation under the Linux Foundation, the same home as MCP.

What is an Agent Skill?

A folder with a SKILL.md file that teaches an agent one job, plus any scripts or reference files it needs. The agent loads it when that job comes up, instead of stuffing every rule into the system prompt.

Do I have to join the AAIF to use skills?

No. Anthropic says you can keep using the standard and contributing without joining. The donation is about who owns the spec, not a new licence tax on SKILL.md files.

Should I install skills from GitHub?

Only if you would run the scripts yourself. A skill can include executable files. Read every line, keep secrets out of the folder, and prefer skills you wrote for your own repo.