An agent skill is a prompt your AI agent loads at the moment a task calls for it. It sits in a library with the rest of your skills, and when work arrives the agent reads the one that matches the job and follows what is written in it.
How does an agent pick a skill?
Ask an agent to write some copy. If there is a copywriting skill in its library it stops before writing a word, reads that skill, and then writes. Invoking a skill is that literal: the task arrives, the agent looks over the library, works out which skill is relevant and which is not, and pulls the text of the winner into its context window.
A library can hold five skills or a hundred. The agent picks.
It picks off the first line or two of each file, the part that says when the skill applies, which is also where this goes wrong. Write that line loosely, or write two skills that both sound like they cover quoting, and the agent either reads the wrong one or reads none and answers out of general knowledge instead. Nothing errors and nothing warns you. You get a fluent reply that skipped your process, and you only catch it by reading the output and noticing which of your steps are missing from it. So write the trigger line about the trigger rather than the topic: “when a customer asks for a quote on a replacement system” beats “quoting”.
What goes in a skill file?
A skill is whatever you put in the file. The ones that work have four parts.
- A line at the top about when to use it. Something like “I will be asking you to do this when I need a blog written.” That line is what tells the agent this is the skill for the job in front of it.
- Examples, especially of the output. Show the finished thing you want back, in the format you want it back in, because the format is the part an agent gets wrong most often.
- A step by step procedure. “Before you write any copy, go and do research.” Write it in order, since the order is exactly what people leave out when they explain a job out loud.
- The tools it can use. If the copy needs a grammar check, say so and point it at a tool connection, a Grammarly MCP.
Is a skill the same as training a model?
Business owners tell us they are documenting everything so they can train the model. They are not training anything. Training a model means hundreds of millions of dollars spent on fine-tuning and retraining, and in practice it means OpenAI, Anthropic, Gemini and the other big labs. Almost nobody else is doing it, and it is out of reach for everyone reading this.
| A skill | Fine-tuning a model | |
|---|---|---|
| Cost | Your time writing the file | Hundreds of millions of dollars |
| Who does it | You | OpenAI, Anthropic, Gemini, the big labs |
| When it applies | At run time, when the task arrives | At training time, before you ever touch the model |
| How you change it | Rewrite the file, the next run uses it | Another training run |
A skill is further instruction on what needs to get done, handed over at run time, which is why you can rewrite it this afternoon and see the difference on the next task.
The reason you need one comes back to the context window. Every time you start with an agent, that window is at zero. It knows nothing. It does not know which quotes need the owner’s sign-off before they reach the customer, or that a no-heat call in January jumps the queue, and it does not remember that you explained both of those yesterday. You tell it again, every time, and a skill is how you tell it without typing. What that buys you is the model handling an after-hours no-heat call the way your best dispatcher would, for that one call, and going back to being a generalist afterwards.
Can I turn my SOPs into AI skills?
Every internal SOP your company has is already a skill, written for a human to follow. The intake script your office manager uses on an after-hours call, the checklist a solar install has to clear before inspection: that is the same material an agent needs, and it is material it cannot get anywhere else, because none of it is on the internet.
When we build agents for clients, we start from the SOPs that already exist, either so the agent can run them or so it has context on what Jane from accounting spends her day doing.
The tell for a skill worth writing is repetition. We hit it with our own coding agents, where pushing code means ten or fifteen checks and someone was retyping them at the end of every session, so it went into a skill. Now it is one invocation and the agent works through the list.
A harness is next, which is everything wrapped around the model rather than fed to it: the tools, the permissions, the loop it runs in.
