Don’t build agents. Build self-improving skills.

A steel fabricator we work with hired us to build agents, a production scheduler and a project manager among them. We talked them into a different list.

What does scheduling the shop actually take? What does a good PM know that a new one doesn’t? Write those down as skills, in how the shop really runs, and any agent can call them. Then it matters a lot less which agent does the calling.

The reason is how agents work. An agent wakes up blind. Every time it’s prompted, it starts with what’s in front of it: its instructions, whatever it can look up, and nothing else. It doesn’t remember the job it scheduled yesterday unless something hands that back. It has never stood in your shop.

The closest thing on a shop floor is the shift change. The night crew didn’t see what happened on days. What they know when they clock in is what’s in the handover log and whatever the day lead remembered to tell them. A good log is short and specific: which job is waiting on galv, which crane is down, who to call if the saw jams. An agent clocks in like that every time it’s called, and most companies hand it an empty log.

Agents need the right context when they’re acting. That’s how they work.

Everything else is in service of that: the skills, the tools, the record of where each job stands.

What is an agent skill?

A skill is a short document an agent reads when the work calls for it. It says how one piece of work is done here: what to look at, in what order, what a number does and doesn’t prove, and who signs. It can carry a script or a reference sheet with it.

Not everything belongs in a skill. At the fabricator, what fits in which week is arithmetic, so we built that as plain software first. The skills are for the choices around it, the ones a good scheduler makes and a new one doesn’t.

Anthropic introduced skills for Claude in October 2025 and published the format as an open standard that December.³ ChatGPT and OpenAI’s Codex read the same kind of file.⁴ So it doesn’t much matter whether your team ends up on Claude or ChatGPT. A skill written for one should carry over to the other, since both read the same open format.

Why not just build the agent?

Because building the agent is turning into the easy part. In September OpenAI launched dots, always-on agents in ChatGPT, each with a computer of its own in the cloud, that learn their owner’s preferences from feedback.¹ In October Anthropic previewed Claude Duet, which checks your connected accounts every few hours and keeps working when you step away.² Next year’s versions of both will be better whether you do anything or not.

Software itself is commoditizing, and agents are software. Of everything that goes into putting AI to work in a company, the agent is the part I’d spend the least on.

What none of them arrives with is how your shop works: what “ready for the floor” means here, when a received quantity is really on the rack, who signs a start date. That’s different in every shop, so no vendor can ship it in a box, and it’s the one part of the setup that’s yours. Helping companies write it down is our business, so weigh that.

Both products remember things, and that’s useful and necessary context. But memory has an agent working from the past: what it happened to see, in whichever account saw it. A skill is what someone who knows the work decided an agent should know, written down where anyone can read it and fix it.

Duet starts with whatever skills are already in your Claude account.² That’s the right design, and it means Anthropic’s newest agent wakes up with none if you haven’t written any.

What makes a skill self-improving?

Self-improving doesn’t mean nobody’s watching. An agent uses the skill, a person catches what it got wrong, and the correction goes into the skill as a line, instead of into a prompt somebody retypes next week. The next agent that calls it, Claude or ChatGPT, starts from the corrected version.

Writing a line down isn’t the end of it. Figure 1 is the version where the skill works. In September we ran four invented cases through the skills we’d written for the fabricator. The conclusions held up, but every answer fell short somewhere, and two said things the evidence didn’t support. In one, the record showed part of a material bill as received, and the agent wrote that the steel had arrived. The skill it had loaded already said a received quantity doesn’t prove the steel is on the rack, and one of our test cases checked for exactly that. It had the rule and said it anyway.

It came up on invented data, in our own testing. A skill needs cases it gets checked against, and a person who still signs the start date. It gets better the way a new estimator does, with someone checking the work every time.

We run our own writing the same way. There’s a skill for drafting in my voice, built from my own posts and calls, and nothing it drafts goes out until I’ve been through it. In October our reviewing agents praised two lines in a draft, and I asked for both to be reworded. The skill now says a reviewer’s approval doesn’t make a line mine, and this essay was drafted from that version.

How do you write your first skill?

Pick the job your best person explains to every new hire. At a fabricator it might be how to tell whether a job is ready for the floor. At a distributor, how to answer “when will it ship?” when the system says one thing and the warehouse says another.

Write it the way they’d explain it, in their order and their words. Ours start as walkthroughs with the people who do the work, not from a procedures manual. Keep it short. The test I use is whether it’s enough to make the call, with nothing in it the agent will trip over.

Give it to whichever agent your team already pays for. When it gets something wrong, fix the line in the skill, not the prompt, and add the miss as a test case so you’ll know if it comes back.

And keep the folder somewhere you own. You’ll probably be using a different agent next year, and the skills should come with you.

What if the part of a business an agent needs to know is a lot smaller than it feels from the inside? I don’t know yet how you tell when you’ve written enough.