Which tasks to turn into an AI agent skill. Describe it once, and it runs without you.

Anything you can describe well enough to repeat becomes a skill an agent can run: the steps, the rules you apply, the way you would handle the awkward cases. What does not cross over is the part that was never a procedure, the direction and the final yes. So the line is not technical. It is whether you can put the work into words.

That is a quieter shift than it sounds, and a bigger one. For most of working life the question was how much you could personally do in a day. The ceiling was your own hours. Now the question underneath it has changed: not how much you can do, but how much of you can be taught to something that does not get tired and does not stop.

Sort your own week and the line draws itself. The describable side is most of it: turning a sales call into a follow-up, formatting the weekly report, drafting the launch post, cutting a campaign into its variants, triaging the inbox, updating the deal notes, answering the question a customer asks for the hundredth time. Each one is a procedure you could write down, so each one can be handed to a skill.

Deciding what is worth making, and whether the result is good enough to carry your name, is not a procedure. It stays with you. The agent runs the volume. One person sets the standard it runs to.

The real work is the same boring back office every business has: reporting, follow-ups, inbox triage, CRM updates, drafting and reformatting, the answers to questions buyers keep asking. People automate it by describing the work once, as a reusable skill the agent loads every time the job comes up, instead of re-explaining it in a fresh chat. The difference between a clever answer and automated work is repeatability, and a skill is where the repeatability lives.

This is now literal, not a metaphor. A skill is a folder of instructions, scripts, and resources an agent loads when it needs them, published as an open standard so the same description runs across different tools.¹ You write down how the work is done, the agent picks up that folder when the work appears, and it runs. Describe once. Run again, and again.

Notice what that does to what multiplies you. A prompt is a one-off; the value evaporates when the chat closes. A skill is durable, because the thinking was captured in a form that can be run, audited, and improved. The move from text generation to real work is exactly this move: from asking for an answer to encoding the method that produces it.

A prompt is a one-off. A skill is the method, written down in a form that runs.

Encode the thing you do daily, do well, and are tired of doing by hand. That overlap is your first skill. The daily part makes it worth the effort, doing it well means you can describe it correctly, and being tired of it is the signal that the work is rote enough to hand off. Start anywhere else and you automate something rare, or something you cannot yet describe.

The reason “do it well” matters is the part people skip. You can only describe a task accurately if you already have the judgment for it, including the edge cases that trip up a beginner. Encoding a task you do badly just scales the mistake. So the best first skill is usually the boring expert task: the one so familiar you forgot it was ever hard, the one your own hands could do half-asleep. That familiarity is exactly what makes it describable.

Yes, and the honest version is more useful than the hype. McKinsey surveyed 1,993 people across 105 nations in mid-2025. It found 62 percent at least experimenting with AI agents, 23 percent scaling one in at least a single function, and no more than 10 percent scaling agents inside any given function.²

Sit with the distance between 62 and 23. Experimenting is easy, because experimenting is a person watching. What thins out is the described method underneath, and the gap is the standard, not the tooling. The companies getting work out of agents are the ones who described the work carefully enough to run.

Start with the example you can point at. A skill has a definition now: a folder of instructions, scripts, and resources an agent discovers and loads, composable with other skills and portable across apps, the API, and an agent SDK.¹ That is the whole mechanism behind describing the work once. The describing is not a figure of speech. It is the file.

We can say the rest first-hand, because the studio runs on it. This article was produced by a journal skill: a described method for researching a question, drafting the piece, and laying it out, run by an agent. The carousels come off a separate carousel skill, the videos off a motion skill. Each is a procedure a named person wrote, the agent runs the volume, and the same person reads and signs every line before it goes out. The proof we trust most is not a benchmark. It is our own output.

None of this says the agents are autonomous, or that they replace the person who set the standard. It says something narrower and more durable. The describable work is now runnable, which is exactly why the part you cannot describe is worth more than ever.

Because a skill has to carry your edge cases, or it is just a prompt with extra steps. A prompt gets you the average competent answer. A skill is the place you put the specific judgment, the “never do this,” the “in this case, do that,” the rule you learned the hard way. That captured judgment is the difference between text and work.

Look at the same job two ways. As a prompt it is a wish: make it good, keep it on brand. As a skill it is the actual rules you would apply if you did it yourself, written down where the agent can run them every time. The brand was never a vibe. It was a stack of decisions, and a skill is where those decisions finally live in a form that runs.

This is the next move past simply getting the thinking out of your head. Extraction is step one: capture what only you know. The skill is step two and the durable one, because it runs that extracted judgment again and again, not once. A transcript is a memory. A skill is an asset.

The brand was never a vibe. It was a stack of decisions, and a skill is where those decisions finally run.

As many as you can still stand behind, and not one more. The agents remove the limit on how much gets produced. They do not remove the limit on how much one person can answer for, and the second limit is the real one. A skill does not scale a person. It scales the standard that person set.

This is why the answer to “how many agents can one founder run” is not a headcount, it is a question about signatures. The measured picture agrees: in a 2026 study of agents running in production, 74 percent depend primarily on human evaluation and 68 percent execute at most 10 steps before a person intervenes.³

In practice that tends to be a handful, not a swarm, because the binding constraint is your own read time, not the agents. You can direct as many skills as you can still read the output of and put your name on. Past that point you are not running skills, you are just hoping, and the work quietly drifts to the same average as everyone else's.

You can run as many skills as you can still read the output of and stand behind.

So the practical path is small and repeatable. Find the first skill, describe it well, run it, read every result, and only add the next one once the last is something you would sign without flinching.

The questions founders ask most about turning their own work into skills an agent runs, answered straight.

Partly yes, and the exception is the whole point. A skill is a described workflow, so the skeptic is right about the mechanism. What the comparison misses is the author. A workflow has no taste and no name on it. A skill carries both, because a person decided what good looks like and signs the output.

That is also the answer to the fear underneath the question, the one about non-deterministic systems in real production. You do not make an agent safe by pretending it is deterministic. You make the output safe by keeping a named human on the read, so nothing reaches a customer that a person did not stand behind. The agent supplies the volume. The person supplies the standard.

So the reframe holds. What multiplies you is no longer the size of your own day. It is how much of your judgment you can describe well enough to run, and how much of the result you are still willing to sign. Describe more of yourself, keep your name on the work, and the work stays yours. That is the whole craft now.