Skills: how agents learn your way of doing things

Promptthis time only

Skill

procedureconstraintsexampleschecks

In plain terms

Every serious company runs on standard operating procedures. A new employee might have brilliant instincts, but you still hand them the operations manual — because “how we do things here” shouldn’t depend on who you ask or what mood they’re in.

A prompt is a sticky note: “please do it this way, this time.” A skill is a page in the manual: written once, versioned, and pulled out automatically whenever that kind of work comes up. The difference sounds small and is everything — sticky notes get lost, reworded, and skipped under deadline. The manual doesn’t.

The anatomy

A well-built skill has four parts. The procedure: the steps, in order, with the decisions called out. The constraints: what must always be true and what must never happen — the non-negotiables. The examples: what good output looks like, because models follow examples better than adjectives. The checks: how to verify the work followed the skill — ideally machine-checkable.

It’s worth placing skills against the alternatives. Prompting is flexible but evaporates — nothing persists, nothing is enforced. Fine-tuning bakes tendencies into the model itself — expensive, slow to change, and probabilistic: the model leans toward your patterns, with no guarantee on any given output. Skills sit in the sweet spot: cheap to write, instant to update, versioned like code — and when paired with checks, enforceable. We put it this way on our library page: training shapes tendencies; encoding creates guarantees.

Skills have failure modes too, and honesty requires naming them: skills rot when the codebase moves and the manual doesn’t, and skills conflict when two of them claim the same territory. Treat the skill library like code — owned, reviewed, versioned — or it decays like documentation.

What good looks like

The question for your own team: are your engineering standards written anywhere an agent could actually execute against? Not tribal knowledge, not a stale wiki — structured, current, checkable procedure. Whatever isn’t encoded is being re-invented per session, by people and agents alike.

Where Cortex stands

This is the episode closest to home. The Kinetic Library is a skill library with scar tissue: fourteen years of production failures turned into named, numbered standards. And in Cortex, skills come with teeth — validation contracts check the work against the standard, so a playbook isn’t advice the agent may consider; it’s a bar the work has to clear.

Questions this paper answers

What is an agent skill? A reusable, structured instruction set for a class of task — procedure, constraints, examples, and checks — that an agent loads automatically when that work comes up. Written once, versioned, applied consistently.

How are skills different from prompts? A prompt instructs once and evaporates; a skill persists, is versioned, and can be enforced with checks. Prompting is a sticky note; a skill is the operations manual.

Can an AI agent follow our internal coding standards? Yes — if the standards are encoded as structured, checkable skills rather than tribal knowledge. The reliability comes from pairing each skill with verification, not from asking nicely.