A company that runs on agents lives or dies by what its agents load before they act. Everything downstream, the email that gets sent, the model that gets picked, the research that gets trusted, is only as good as the procedure the agent reached for first. Prompts are wishes. Skills are procedure.
Here is the difference, concretely. You can tell an agent to verify an email address before sending anything to it. That is a wish. It holds until the context window resets, until a new agent joins the fleet, until the same agent three tasks later has quietly forgotten the sentence nobody enforced. Prose guidance decays. It always decays, because it depends on the agent choosing to remember it at the exact moment it matters.
Or you give the agent a skill. Not a sentence, a file it loads as operating procedure. My email-finding skill is a few hundred lines: it fires automatically the moment an agent needs someone's address, walks a fixed escalation ladder (free lookups first, paid tools only if those fail, ask a human before guessing), and ends on a rule that cannot be talked around. Never ship an unverified address. The agent follows it whether or not anyone is watching, because following it is the only path through.
That is the whole shift. A wish is advice the agent may take. A skill is a road the agent has to walk.
It matters more as the models get better, not less. A weaker agent needed the hand-holding of a long prompt. A stronger one improvises confidently past a vague instruction, produces a competent, plausible answer, and is wrong in the exact way the wish was trying to prevent. The smarter the agent, the more the guardrail has to be structural, because a capable reasoner will talk its way around anything softer.
Four things, in my experience.
A trigger, so it loads itself. The best skills fire on conditions, not on being remembered. "Whenever you are about to find, verify, or prospect an email" is a trigger. It means the discipline arrives on its own, at the moment of the decision, instead of waiting for the agent to recall that it exists.
Decision rules, so judgment is not improvised. The escalation ladder, the order of operations, the thresholds. This is the part that used to live in a senior person's head and now lives in a file every agent reads the same way.
Hard gates, so the important line cannot be crossed. A cap that is only a comment is not a cap. "Never ship unverified" has to be a wall, not a preference, or it erodes the first time it is inconvenient.
Anti-patterns, drawn from real failures. Every skill I keep carries a list of the specific ways the work went wrong before, written down so the mistake is made once and then structurally impossible to repeat. This is where a skill stops being generic advice and becomes ours.
A human operator's judgment leaves when they leave. You can write some of it into a handover doc on the way out, and the doc will rot within a quarter. A skill does the opposite. Every lesson the company learns once, it learns permanently: someone hits a failure, the fix goes into the skill, and from that moment every agent that will ever run inherits it. The layer accrues. It is the rare asset that gets sharper as the team gets smaller.
This is what I mean, mechanically, when I say keep a copy of the operator. The operator is not a person you retain. It is a stack of skills that encodes how the company thinks, and it stays after any particular engagement ends.
At my last company the fleet was seventy-plus scheduled agents running across eighteen countries and eleven languages. They did not each learn the job independently. They inherited the same skills, the same ladders, the same hard gates, the same list of mistakes already made. Two people could run that surface area because the judgment was written down once and loaded everywhere.
The collection is open source. I pulled the general skills, the model-picking discipline, the email-finding ladder, the research hygiene, the anti-hallucination checks, out of production and put them in a repo called operator-skills. It is a fraction of the layer, but it shows the shape. And it is the same layer I install first at every company I work with, because a company's agents are exactly as good as what they load before they act.
I write short memos on running companies AI-native. The skills that teach my agents judgment are on GitHub; the practice that installs them is at raianpollock.com.