What Is a Compiled Skill, and Why It Beats Recording a Human
TL;DR
A recording captures coordinates and timings, which die with the session that produced them. A compiled skill captures intent: what the person was doing and why, written as steps a model can follow on a page that has since changed. Skills should carry a track record, be handed out only once until they prove themselves, and retire when they start failing, so a fleet's knowledge improves rather than rots.
A recording says click at 812,344. A skill says dismiss the cookie banner, then open Sign in. The first breaks the next time the button moves. The second survives a redesign, because intent does.
Recording versus compiling
| A recording | A compiled skill | |
|---|---|---|
| What it stores | Coordinates, keystrokes, timings | Steps written as intent |
| Survives a redesign | No | Usually |
| Readable by a person | Barely | Yes |
| Followable by a model | No | Yes |
| Contains typed secrets | Often | Never |
How a skill gets written
A person fixes a stuck agent in a live browser. The system keeps one screenshot per action they take and the shape of each action, but never the characters typed. A vision model then reads those moments and writes the procedure: not what pixels were clicked, but what the person was accomplishing at each step. If the recording does not support a clear procedure, it declines to write one rather than inventing it, because a confident wrong skill is worse than none.
Why a skill needs a track record
A freshly compiled skill has never been tried. Hand it out once; if the agent that followed it succeeds, it earns trust and is offered again. If agents following it keep failing, it retires itself. The library that results gets better with use instead of accumulating stale procedures, which is the failure mode of every shared-macro system ever built.
# Skill: dismiss the consent banner on vendor.example, then open Sign in
Use when: the page shows a consent dialog covering the header.
1. If a consent banner is visible, choose the option that accepts or closes it.
2. In the header, open "Sign in".
3. If a login form appears, stop and ask for a human; credentials are not part of this skill.That third line is the important one. A skill knows the edge of what it should do, and hands the rest back to a person.
Questions people ask
What is a skill in the context of AI agents?
A procedure written as intent, compiled from a human's fix, that an agent follows the next time it meets the same wall. It describes what to accomplish at each step, not which pixels to click.
Why not just record the human and replay it?
A recording stores coordinates and timings that break the moment the page changes. A skill stores intent, which survives a redesign, and never contains the characters a person typed.
How does an agent know which skill to use?
It asks before attempting a task, describing the task in plain words, and receives the skills whose record shows they work for that wall on that site.
What happens when a skill stops working?
Agents report whether it worked. A skill whose failures outnumber its successes is no longer offered, so the library improves with use rather than rotting.
DoubleOh is the reliability layer for AI agents. When one gets stuck, a person fixes it once in a live browser, and the fix becomes a skill the whole fleet follows from then on.
Start free