ARTICLE · CONTROLS FOR AI WORK · PART 2 OF 6 · Work instruction at the station

Skills: the last thing read before the action

Say you need to print a document. You look up the procedure for printing, read it, and then print. Every time you print, you read it again first. That procedure gets a special place: it’s the last thing read before the action, and nothing said earlier in the conversation gets to argue it down. That’s what a skill is for an AI agent, and it’s the same thing a work instruction at a station has always been for a person.

The same questions apply to a sales pipeline, a purchase order or a month-end close: what stays fixed, what needs judgment, and who decides?

What the AI field has found

One of the major AI labs published two pieces that explain why this works.

The first is on what they call context engineering. A model has a limited “attention budget”: the more text you load into a conversation, the less attention each piece gets, and its ability to recall details declines (they call it “context rot”). Their guiding rule is that good context engineering “means finding the smallest possible set of high-signal tokens,” meaning the least text that does the job. They recommend loading information “just in time,” when the task needs it, not all at the start.

The second is on Agent Skills. An AI agent starts with only the name and a one-line description of each skill it has. The full instructions load only when the AI agent decides a skill applies to the job in front of it.

Other research backs this up from the opposite direction. The IFScale study gave AI models more and more instructions at once. Rule-following dropped as the list grew; even the best models followed only about 68% of 500. Models also favored the instructions near the top of the list. One giant list, loaded all at once, is a weak way to give instructions.

From the shop floor: the work instruction at the station

I’ve written procedures, setup sheets and work instructions for most of my career. In one plant, we set up the method the whole plant ran on, and the most important thing we learned was to keep only the documents people actually needed, where they needed them.

A good work instruction isn’t the plant’s entire quality manual taped to the machine. It’s the steps for this job, at this station, read when the job starts. The operator doesn’t need the molding procedures while packing boxes. Give them everything and they’ll read nothing.

A skill is a work instruction for an AI agent:

  • One job. Printing, or saving a file, or building a cost sheet. Not everything the AI agent might ever do.
  • Read at the moment of the action. Not at the start of the conversation, where forty exchanges later it’s a small voice in a long chat.
  • Read again every time. The tenth print gets the same instructions as the first.
  • Specific enough to check. “Print on letter size, check the page count, save a copy to the project folder” can be followed and checked. “Be careful and thorough” can’t.

Here’s how the A.N. Creation Office uses it. Its Project Manager Bot has about two dozen principles. It doesn’t load them all:

b-project-manager reads this list at the start of every conversation, and opens a principle’s file when it applies.

The list is short: one line per principle, saying when it applies. The full text opens only when the situation calls for it.

The parking lot full of cones

I learned the other side of this the hard way.

Early on, working with AI coding tools, we kept one instructions file that the AI agent read at the start of every session. Every time something went wrong, we added a line. Each line made sense on its own. Nobody planned the whole thing. Over time it got more and more constrained, and the AI agent got worse at its job, not better.

It was like driving into a parking lot after someone has thrown a pile of traffic cones everywhere and told you not to hit any of them. Each cone is there for a reason. Together, they make a mess that’s hard to get through.

A skill is the opposite of a cone. It’s one planned set of instructions for one job, brought out when that job comes up and put away after.

Plan the revert

Skills have one more thing in common with any system you build: the world changes under them.

Years ago I built a database to track the pad-printing dies in a plant: where each one was, which press it went to, who had it checked out. It worked well for a year and a half. Then a software update came through, the forms broke, and the totals doubled: thirteen thousand by hand, twenty-six thousand on the screen. The vendor never fixed it. I didn’t have a week to debug it, so I exported everything back to a spreadsheet in a few hours.

AI tools do the same thing. I had a set of skills that worked well on one model version, bogged down on the next, and nearly stopped working on the one after. Many of them existed to make up for the AI agent forgetting things, and the newer models didn’t forget. The fix was to strip them out.

So build skills the way you’d build anything a plant depends on:

  • Keep them short, so they’re easy to read and easy to change.
  • Know why each one exists. If the model now does it without being told, cut it.
  • Make the revert cheap. If a model update starts fighting your skills, you should be able to take them out in an afternoon.

What breaks a skill

  • It never loads. If the AI agent doesn’t recognize “this is a printing job,” the skill never comes up. Clear names and descriptions matter, and so does the playbook saying “this step uses this skill” (Part 3).
  • There’s talk between loading and acting. If the skill loads and ten more exchanges happen before the print, it’s no longer the last thing read. Load it right before the action.
  • It’s too long. A skill that tries to cover everything dilutes itself.
  • Two skills disagree. If two skills format the same number differently, the AI agent picks one. Decide which rule wins (Part 4).

Try this

Take one task you give an AI agent over and over. Write the steps on one page: what to do, in order, and what done looks like. Before each time you do the task, paste that page in first. Compare the results to what you got from one long set of instructions at the start of the chat.

Sources

These pages sit with the rest of the Advance Now family. The hotbotspot tool library writes the methods up for AI. A.N. Creation Office is where the public AI agents run as open experiments.