Built AI-first · Chapter 4 of 9
Context, Skills, Routines. What is true, how we do a thing, and when it runs. Every AI-first organisation we have built or examined comes down to these three, whatever they are called locally — and the discipline that makes them work is refusing to let them mix.
This is the least exciting chapter in the series and the one that decides whether any of the others hold. Mixed-together context, method and schedule is the swamp that every stalled AI programme is standing in.
The claim
A fact changes when the world changes — a price rises, a policy is superseded, a client's scope moves. A method changes when you get better at it, which is on your schedule and nobody else's. A cadence changes when the calendar or the organisation changes, which is rarer than both.
Put all three in one prompt, one document or one assistant configuration and every change to any of them risks the other two. You cannot improve the method without stepping around the facts. You cannot update a fact without hunting for every place it was embedded. And you cannot tell, afterwards, whether a bad result came from a stale fact, a weak method or a run that never happened.
Keeping them apart is not tidiness. It is the only way to diagnose a failure after it happens.
The three
File one
What is true here.
Facts, settled decisions, constraints and voice — the library from chapter three. Read before anything is drafted, cited when anything is produced.
File two
How we do a particular thing, well, every time.
A written procedure: what it takes in, the steps, the checks it must pass, and what it refuses to do. Not a prompt — a method, of the kind you would hand a competent new colleague.
File three
When it runs, who owns it, and who is owed the result.
The standing cadence. Weekly, monthly, on a trigger — plus the part everyone omits: what it means when the routine did not run.
The test
Take any sentence in your current prompts, playbooks or assistant configuration and ask what kind of thing could go wrong with it.
| If it could be… | It belongs in… |
|---|---|
| True or false | Context. "Our standard payment term is thirty days." Either it is or it is not, and the answer has a date on it. |
| Done well or badly | Skills. "Write the renewal note." There is no true or false here — there is competent and incompetent, and the difference is written down as steps and checks. |
| On time or late | Routines. "Every Friday, before noon." Nothing about this is a fact or a method. It is a commitment with an owner and a recipient. |
Run it across one existing prompt and the usual result is three or four sentences that each belong somewhere different, currently welded together. Separating them is often the first hour of real work in an AI programme, and it is the hour that makes the next fifty tractable.
Why it compounds
● Code red · welded together
A better way of writing the renewal note is trapped inside the prompt where it was discovered, alongside that prompt's copy of the pricing and its own idea of when to run. Nobody else benefits. When pricing changes, six prompts contain the old number and five of them get updated.
● Code green · kept apart
The method improves once and every agent calling it improves. The price changes in one entry and everything downstream is immediately correct. The cadence moves without touching either. This is the structural reason an AI-first organisation gets better over time and a retrofitted one plateaus.
It also makes the next two chapters possible. An agent, in the sense we mean it, is a named job that holds skills and reads context on a routine — so the three files are the parts an agent is assembled from. And autonomy, in chapter eight, is granted to a routine rather than to an agent, which you cannot do at all if the routine is not a separable thing.
In your world
None of this is new to any of these organisations. All four already run on context, method and cadence; what is missing is that any of it is written where something other than a person can read it.
The policy guide is context, the standard operating guideline is a skill, and the drill and testing calendar is a routine. The vocabulary already exists — it is one of the few sectors where these three are genuinely distinct documents, which is a real head start.
Municipal code and adopted resolutions are context, the clerk's procedures are skills, and the council cycle is the most unforgiving routine in local government. The gap is usually that the procedures live with the clerk rather than on paper.
The price book is context and often the only one of the three written down. How you scope a job is a skill living in the owner's head, and the follow-up cadence is a routine that mostly does not exist — which is the same finding as chapter two.
Scope and commitments are context, your delivery method is a skill you probably believe is unwriteable, and the weekly client rhythm is a routine. Writing the method down is the hard one and the one that makes the practice sellable.
What we will not claim
Where this sits
Context, skills and routines are the materials. Chapter five is about the two roles that turn a drawer of them into a system that compounds — and the fact that most agent teams are missing both.