What it is, and why loading matters more than the file
Everything an agent knows in a conversation has to fit in its context, and everything in its context competes for its attention. Pasting your house style, your review checklist and your deployment procedure into every session is how a model ends up ignoring all three: the instruction that matters for this task is buried among twenty that do not.
A skill inverts that. The agent sees only the name and a one-line description of when to use it — a few dozen words — and reads the full procedure at the moment it decides the job applies. What is not relevant costs nothing, so a hundred skills can exist without a hundred skills' worth of noise. This is usually called progressive disclosure, and it is the whole mechanism.
The format is deliberately dull: a Markdown file, optional scripts and reference files beside it, a folder you can read and put in version control. Being ordinary files is what makes them reviewable, diffable and shareable at all.
It is a written-down way of doing one job, which the agent picks up when that job appears.
How it differs from a prompt, a tool and an MCP server
A prompt is what you say this time. A skill is what is true every time this job comes up, written once and applied without anybody remembering to paste it.
A tool is a capability — something the agent can do. A skill is judgement about doing it: which order, which checks, what counts as finished, what to refuse.
An MCP server gives an agent reach into a system. A skill gives it competence at the work. They are complements, and the failure mode of having one without the other is easy to picture: an agent connected to your database with no skill will write queries that run and answer the wrong question.
What separates a good skill from a long one
- A description that says when to use it, not what it is. The description is the only part an agent reads before deciding, so "use when reviewing a pull request in this repository" beats "code review helper".
- One job. A skill that covers reviewing, releasing and rolling back is loaded for all three and precise for none.
- Procedure over prose. Steps, checks and worked examples survive being followed; an essay about quality does not.
- The refusals in writing. What not to do, and when to stop and ask, is the part a general model gets wrong most often and the part your team knows.
What to be careful about
- A skill is instructions an agent will follow. Installing one from a stranger is trusting its author with whatever the agent can do, and the text is not audited by anyone in between.
- Skills drift like documentation. One that describes last quarter's process will be followed anyway, confidently, which is worse than having none.
- More is not better. Dozens of overlapping descriptions make the choice harder, and the failure looks like the agent picking a plausible neighbour rather than the right one.
- A skill cannot fix a model that lacks the capability. It directs competence; it does not supply it.
Why it appears in a glossary about visibility
Because it is the clearest case of writing addressed to a machine reader that is still writing. A skill is judged on exactly the qualities a page needs to be quotable by a retrieval system: does the first line say what this is for, is the useful part near the top, is a step something you can lift out and act on without the surrounding paragraph.
There is a practical connection too. The work your buyers do with agents is increasingly done through skills somebody published, and what those skills say about a category is another place your product is described without you. We keep a public directory of them for that reason.
Frequently asked questions
Is a skill the same thing as a custom GPT or an assistant persona?
No. A persona is a standing instruction that shapes every reply in a session. A skill is loaded for one kind of work and ignored otherwise, which is why an agent can hold many skills and only one persona without them fighting.
Do I need to be a programmer to write one?
No. The core of a skill is a Markdown file describing how a job is done properly — the thing your best practitioner would tell a new colleague. Scripts are optional and many useful skills have none.
Where do skills come from?
You write them for your own work, or you install ones other people published. Anthropic publishes an open set, and there are community directories; ours lists what is public, with attribution, and links back to the source rather than copying it.
Do skills work with any AI assistant?
The format was introduced for Claude and is supported across Anthropic's surfaces, and other agent frameworks have adopted similar loading conventions. Portability is not guaranteed the way a protocol guarantees it — a skill is a document plus a convention for finding it, not a standard both ends implement.
Asked out loud
spoken, not typedThe same term in the words somebody uses speaking to an assistant rather than typing into a box — written from the situation, which is why each one carries the situation it came from.
That is what a skill is for. Take the instructions that are true every time, put them in a file with a description of when they apply, and the agent loads them at the moment the job comes up instead of carrying them through every conversation and losing them among the rest.
A server gives the agent reach — things it can read and do; a skill gives it know-how — how to do a particular job well. Reach without know-how gets you confident work on the wrong question, and know-how without reach gets you good advice and no action.
Open it and read it, because it is a text file and reading it is the audit. You are trusting the author with whatever the agent can already do, so the questions are who wrote it, whether it asks the agent to run scripts, and whether anything in it tells the agent to skip a confirmation.