Vol.083 — SKILLS: It has twenty-eight of them and uses none

Vol.083 — Multi to Single Agent Journey Issue 6 SKILLS

Date: 2026-10-23 | Newsletter · Multi to Single Agent Journey #6 of 10


A “skill”, in the tool I use, is a packaged procedure — a folder of instructions the AI staff member can run with a slash command, built once so the same job comes out the same way every time. I had built seventeen of them by September. I was pleased with that number. This issue is about what I found when I counted how many were actually being used, who could see them, and why an AI staff member with twenty-eight skills available behaved, most days, as if it had none.

Twenty-eight skills, six staff, and nobody told me they were shared

The first finding came on 18 September, in the paper that took the six-staff structure apart (that story is #3). Skills and connectors in this tool are set at the account level, not per project. So every one of my six AI staff had always been able to see all twenty-eight skills — my seventeen plus the ones that came with the tool — and every connector.

That changes what a skill is for. If every staff member can run every skill, a skill is not a way of giving one staff member a capability. It is a way of making sure a job is done the same way whoever does it — which, once there is one agent, means the same way every time. That is a smaller and more useful idea than the one I had.

“Skills are used only when named” — and so they were never used

The second finding was about my own rule. Until 20 September the standing rule was that a skill is used only when I name it. I had written that out of caution: I did not want the staff running a heavy procedure when I had asked a light question. The effect was that the skills sat there. I would ask for a teaser or a buyer list in ordinary words, the staff would build it by hand from scratch, and the skill I had spent an afternoon making would not be touched, because I had not said its name.

On 20 September I replaced the rule with its opposite: detect and use. When a registered skill fits the work, use it, do not wait to be told — but say the skill’s name and what it will produce before running it, in one line, so that I can trim the scope or stop it in one word. “Building the teaser with noname-teaser — pptx now, screened buyer PDF on finalisation.” I made that call knowing it would produce the occasional wrong use. At the volume we work at, I would rather catch a miss than be asked a question on every task, and overuse is how the agent learns where the boundary actually is. That reasoning is recorded next to the rule as provisional, because I expect to revisit it when the volume changes.

Seventeen became nine, by one test

Then came the cut. On 23 September, when the rules for the new single agent were compiled from the record, the skills were compiled too, and the test was one sentence: a skill exists because a lane names it. If no rule for any kind of work says “this job is done with this skill”, the skill has no reason to be registered.

Seventeen of my own became nine. Eight of the nine are each named in a ruling; the ninth I kept by my own call. The eight removed had something in common: none of them was named anywhere. Two of them had actually been run — a web audit once in August, a usability module with files dated June — and I cut them anyway. Keeping a skill because it once ran is how a list grows past what anyone can hold in mind. A skill I need again can be re-created with the agent in an afternoon from the record of what it did; a skill that sits unused costs attention every time the list is read, and the agent reads the list every session.

One honest footnote. Removing a skill is done in the tool’s own settings, by hand — the files the agent can see are a read-only cache. As I write this, the eight are still registered, because I have not yet gone in and removed them. The list the agent reads at the start of every session is still twenty-eight long. Cutting them on paper was the easy half.

“Acceptance test comes after the actual uses”

The part of this I think is worth most to other people is the method that came out two days later, on 25 September, because it answers the question I had never asked: when is a skill finished?

The agent proposed a method with an acceptance test as the last step of the build. I rejected the placement. “Acceptance test usually comes after the active tests of actual uses of slash skills. Not during the build.” So a skill is born in the status Prototype — the build proves nothing — and it earns the question of promotion only after its second real invocation on real work. The agent counts the invocations; I decide the promotion, because the count is mechanical and the judgement of whether the output was worth having is mine. A skill that performed beautifully in the session that made it, and was never called again, stays Prototype forever. That is the intended reading, not a side effect. Retired skills come out of the tool, and every version stays in my folder.

The agent also proposed a trigger rule for building a skill — something that would name good session work before it evaporated. I declined. Either I say “let’s build a skill” at the start, or, after a series of conversations where the quality is visible and the repetition is expected, I ask it to build one from what the session did. On building a skill, I rule. What that puts on the agent is legibility: a session has to be recorded well enough that “make this a skill” is a real instruction and not an invitation to reconstruct the afternoon from memory. Which, if you have read #2 and #3, is the same lesson again from a different door.

Three small things I was asked to rule on, and what I said

Three items in that review I refused to rule on until they were explained to me, and I think the explanations are useful. A refusal condition is what a skill does when the ground gives way, so it stops instead of guessing — my teaser skill refuses to build a deck without a hero image; my CIM skill stamps SOURCE PENDING rather than write an unsourced block. I ruled that each skill writes its own refusals: a generic inherited “stop if unsure” is satisfied generically, and the refusals that hold are the ones written by someone who knew what this particular job breaks on. The past-versions discipline is just file hygiene applied to skills — replace, never stack, never lose what was replaced — and I moved it to a quarterly clean-up rather than a per-change job, so ordinary days stay under one rule instead of two. And helper scripts — the one-job converters and publishing calls the agent writes and throws away — are not skills, not versioned, not reviewed. The agent noted against that ruling that those scripts had published every page of our internal site, so the site is reproducible only while someone will rewrite them. I let the ruling stand. It is on record, which is where it should be if it ever matters.

My Learning: It’s a Good Practice to review your skills periodically.

If you have built skills and your AI staff never seems to use them, I would check the rule before the skills. Mine said “only when named”, which in practice meant never. If you have built a lot of them, I would run the one-sentence test: is this skill named by any rule for any job? If not, it is a demo, not a tool, and the list it sits on is read every single session. And if you are wondering whether a skill is finished — it is not finished when you build it. It is finished after the second time it does real work and you were glad it did.