Out of the box, a model can explain how to do your invoicing procedure. It cannot actually run it - pull the record, apply your rules, produce the document the same way every time. A skill is what closes that gap.
What a skill is
A skill is a packaged, repeatable procedure a model can call on demand - a defined set of steps, wired to your tools, that produces the same correct result every time. Where a plain chat produces a fresh improvisation each time, a skill runs a known procedure reliably.
Skills vs. prompts vs. plain chat
A prompt is an instruction for one response. A skill is a reusable capability. You would not re-explain your entire refund process every time you wanted it done; you would define it once as a skill and call it whenever needed - by a person or by an agent.
What skills let a model actually do
- Reach into your systems - read and, within limits you set, safely write records.
- Run a defined procedure the same way every time, not a new guess.
- Produce the documents, summaries, and replies your team makes over and over, in your format.
What makes a good skill
The best skills are defined (the steps are written down), scoped (they can only do what you allow, with guardrails and logging), and reliable (the hundredth run matches the first). Once built, skills become building blocks - an agent is far more capable when it has a library of skills to call. Explore AI skills.
If you can write down the steps of a task, a skill can probably run them. Describe a repeatable procedure and we will tell you honestly whether a skill fits.