A company that runs on agents lives or dies by what those agents load before they act. Everything downstream, the email that goes out, the model that gets picked, is only as good as the procedure the agent reached for first. Prompts are wishes. A skill leaves the agent no choice.
Concretely. You can tell an agent to verify an email address before it sends anything to it. That holds until the context resets or a new agent joins the fleet, and then the sentence nobody enforced is just gone. Prose decays. It depends on the agent choosing to remember the rule at the exact moment the rule matters, and that moment is usually the inconvenient one.
A skill is a file the agent loads as operating procedure. Mine for email-finding runs a few hundred lines. It fires on its own the moment any agent needs someone's address, walks a fixed ladder from free lookups to paid tools to asking a human, and ends on a line with no way around it: never ship an unverified address. The agent follows that whether or not anyone's watching, because following it is the only path the tooling leaves open.
That's the entire shift, and it sounds small until you watch it hold at 2am with nobody in the loop.
It matters more as the models get better. A weak agent needed the hand-holding anyway. A strong one improvises confidently past a vague instruction and produces something plausible, wrong in the exact way the instruction was trying to prevent. A capable reasoner will talk its way around anything softer than a wall.
Four things, in my experience.
A trigger, so it loads itself. The best skills fire on conditions instead of on being remembered. "Whenever you're about to find or verify an email address" is a trigger. The discipline shows up on its own at the moment of the decision.
Decision rules, so judgment isn't improvised. The escalation ladder and the thresholds, 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 line that matters holds. "Never ship unverified" can't be a preference, because a preference erodes the first afternoon it's inconvenient. So the line lives in the tooling, a check the agent can't complete the task without passing.
Anti-patterns, drawn from real failures. Every skill I keep carries the specific ways this work went wrong before. A mistake gets made once, written down, and becomes structurally hard to repeat. This is where a skill stops being generic advice and becomes ours.
A human operator's judgment leaves in the elevator with them. Some of it survives into a handover doc, and the doc rots within a quarter. A skill moves the other direction. Someone hits a failure, the fix lands in the file, and every agent that will ever run inherits it from that day forward. The layer keeps accruing even as the team gets smaller, which is a strange property for an asset, and the one I lean on most.
It's also the mechanical version of a promise on my front page, that what I build keeps working after I leave. What stays behind is the stack of files, the encoded version of how the company thinks.
At Flip Education the fleet reached seventy-plus scheduled agents working across eighteen countries and eleven languages, run by two people. None of them learned the job independently. They inherited the same skills and the same list of mistakes already made.
The collection is open source now. I pulled the general-purpose pieces out of production, the model-picking discipline, the email ladder, the anti-hallucination checks, and put them in a repo called operator-skills. It's a fraction of the layer, enough to show the shape. It's also the layer I install first at every company I work with.
The skills my agents load are on GitHub. Installing them at other companies is the practice.