Agent skills are saved sets of instructions, sometimes with scripts or reference files, that an AI agent loads and follows whenever a matching task comes up. They give an agent a consistent method, so the tenth weekly report is put together the same way as the first.
Skills sit alongside memory rather than replacing it. Memory holds facts about you, such as your role and your preferences. Skills hold methods for doing a particular job.
What's inside an agent skill?
At minimum, a name, a short description of when to use it, and written instructions. Many skills add examples, reference files or small scripts.
The exact structure depends on the engine. In the open Agent Skills specification, a skill is a folder built around a file called SKILL.md, and Anthropic's Agent Skills documentation explains that the agent sees only each skill's name and description until a task matches.
Look inside a typical skill and you'll find:
A trigger. When should the skill be used? "When I ask for a proposal", say, or "when I paste a new enquiry". In the open format, this lives in the description, and without it the agent has no reason to pick the skill up.
Instructions. The method, written as steps in the order they happen.
Rules and examples. What good output looks like, what to avoid, and the exceptions that matter.
Access. What the skill can read, write or call. Some formats let a skill list the tools it's allowed to use, though the specification marks that as experimental and support varies. More often, a skill can use whatever the agent itself can reach. That's where most of the risk sits, and few people check it.
How is a skill different from a prompt?
A prompt works once, in one conversation. A skill is stored, and the agent picks it up by itself whenever the task fits.
If you've written a good prompt and found yourself retyping it, a skill is the next step. A saved template sits in between: you paste it, adjust the details and run it, which beats retyping but still depends on you remembering it exists. With a skill you ask for the outcome, and the agent recognises the task and selects the method itself. That's why skills keep getting used months later, when templates have long been forgotten in a notes app.
Can an AI agent write its own skills?
Some can. When an agent works out a method that succeeds, it can save the method as a new skill and reuse it next time. Hermes Agent and OpenClaw both do this, for example after a multi-step task or after you correct their approach. Solve a problem together once, and over a few weeks the agent builds up procedures specific to your work.
The same ability can lock in a mistake. If the agent settled on a poor method on a confusing day, that method keeps coming back until someone fixes it.
Engines handle review differently, so don't assume there's an approval step. Hermes Agent saves skills freely by default, with an optional setting that holds every new or changed skill until you approve it. OpenClaw applies changes straight away, but keeps earlier versions so any change can be undone. Both offer automatic checks that look for risky content in agent-written skills.
To keep self-written skills safe:
- Review them. Ask to see what the agent has saved, especially in the first month, and read the instructions as well as the name.
- Make changes reversible. Where approval is available, switch it on for anything important. Where it isn't, make sure you can see what changed and roll it back.
If a platform won't let you inspect what the agent has learned and undo it, treat self-written skills as a risk. In Able°, skills from our library are versioned and checked by us before release, and none of them update without your say-so.
When is a skill worth building?
When the task repeats at least weekly, the method stays the same from one run to the next, and you can describe what a good result looks like, ideally with an example. If the right approach changes every time, a prompt will do. A skill for something you do twice a year mostly adds clutter.
Good early candidates:
- turning meeting notes into actions with owners and dates
- drafting a standard reply to a common type of message
- preparing a weekly summary in a fixed format
- checking a document against a checklist before it goes out
Poor candidates:
- anything involving negotiation or pricing judgement
- one-off research
- tasks where the right answer is whatever the most senior person decides that day
Timing matters too. In your first month, use the agent and notice what you repeat before packaging anything; your first week is about finding one job, and a skill written before you understand the task tends to preserve the wrong method. A good check: if you couldn't explain the method to a new colleague in five minutes, it isn't ready to become a skill.
How do skills affect what your agent can access?
A skill often needs to do something: read a folder, query a system, send a message. Each of those is a permission, so give the narrowest access that still does the job.
- If a skill only reads, it shouldn't be able to write.
- If it drafts for your approval, it shouldn't be able to send.
- If it works on one client's material, it shouldn't see the others.
Access tends to creep up over time. Each new skill can be the reason someone connects another tool, and an agent with thirty skills a year from now may reach far more than anyone intended. Review the list now and then, the way you'd review who can open a shared drive. Skills written by other people need the same care: Anthropic advises using them only from trusted sources.
How AI agent safety works covers the same principle in more depth.
Signs a skill has gone stale
- The output format no longer matches what you need.
- It mentions a person, client or process that has changed.
- It keeps producing the same mistake, because the rule inside it is wrong.
- Nobody remembers who approved it or why it exists.
A stale skill produces outdated work with complete consistency, which makes it harder to spot than a one-off error. Fixing the step inside the skill is the deepest kind of correction, covered in How to correct an AI agent.
Common questions
What is the difference between a skill and a tool?
A tool gives an agent access to something, such as a system it can read from or an action it can take. A skill gives it a method: the steps, rules and examples for doing a particular job well. They often work together, with a skill telling the agent which tool to use, in what order, and what a good result looks like.
What is a SKILL.md file?
It's the main file in a skill folder under the open Agent Skills format. It opens with a short header giving the skill's name and a description of when to use it, followed by plain written instructions. The folder can also hold examples, reference documents and scripts, which the agent reads only when the task needs them.
Can you use skills made by other people?
Often, yes. Many agents can load skills written by others, and shared libraries exist. Treat them like installing software, because a skill can tell your agent to run code or use tools in ways its description doesn't mention. Read the instructions and any scripts before you add one.
Can you switch off or delete a skill?
Usually, yes. You can ask the agent to stop using a skill, remove it or show you its contents, and many engines also store skills as folders that can be deleted directly. If a skill has caused problems, check whether the agent saved a newer version before you assume it's gone.