Where do you start? The first routine you hand over.

In short: The task that annoys you most is usually the worst first pick for an AI operator, because it is unclear and hard to describe. Four criteria — describable "done", weekly repetition, cheap mistakes, narrow access — show which routine actually belongs first.

Anyone who has decided to use an AI operator reaches first for the same task almost every time: the one that eats the most hours and causes the most irritation. That is understandable and usually the wrong start.

The task that annoys you most is often annoying precisely because it is unclear, variable and full of exceptions. It grates because it is poorly defined. And poorly defined work is the hardest thing you can hand to an operator.

You do not pick the first routine by how much it hurts. You pick it by how well you can describe "done".

Four criteria for a first routine

Four properties make a good first role. A routine that meets all four is a good start even if it saves you very little time. The first case exists to prove the mechanism works, not to rescue your calendar.

1. "Done" is describable. You can say how a good result is recognised, precisely enough that another person would reach the same verdict. This is the hardest of the four and the most important. Without a defined standard there is no quality confirmation, and without that you do not have governance, only output.

2. It repeats. Weekly at least, ideally more often. Repetition is what produces evidence: with a task that comes round four times a year it takes years to learn whether quality holds. With a weekly task you know in six weeks.

3. A mistake is cheap. Ask what happens if the result is wrong and nobody notices immediately. Does that cost an afternoon or a customer? For a first role you explicitly want the cheap version, because that is where you learn how your check behaves in practice.

4. Access can be cut narrowly. The work has to be doable from a few clearly bounded sources, and it should not need write access to anything critical. If a task only works with full access to your core system, it is not a first role.

What does not qualify yet

The counter-list is just as useful. Four kinds of work do not belong in the first round, and in three of the four cases that is not permanent.

Work without a definable standard. If "good" at your company means a particular person nods approvingly, the yardstick is missing. That is not an exclusion forever, it is a task that needs a standard first.

One-off work. A relaunch, a migration, a tender. Without repetition no evidence accumulates and the effort of writing the Role Contract never pays back.

Work with a wide blast radius. Anything that reaches customers, money or legal matters directly and unfiltered. That comes later, once the adoption levels rest on measured evidence rather than confidence.

Actual decisions. Anything sitting at Root or Trunk level under the sort by impact. This is the only item on the list that holds permanently: an operator executes, it does not decide the existential questions.

Three of those four exclusions are appointments for later. Only the last one is a boundary.

How to rank your candidates

Take twenty minutes and write five to ten recurring pieces of work underneath each other. Not departments, concrete sequences: the weekly sales report, the answer to standard enquiries, the preparation of the monthly numbers, the follow-up on open quotes.

Then walk the four criteria and give one point per criterion. Whatever lands on top is your first role. It will rarely be the one you thought of first, and that is the actual output of the exercise.

Something second almost always happens too, and it is more useful than the ranking: on several tasks you notice at criterion one that you cannot describe "done" at all. That is not an aside. That is the list of places where your understanding of quality lives in people's heads rather than in a standard, and you can work through it independently of any AI.

Company 0

Our own first role was the weekly content production, and it was not the task consuming the most time. It was the one with the best fit, on all four criteria.

"Done" is describable here because a checklist exists, currently thirteen edit patterns that every text runs against, with FTT (First Time Through) over the top. The work repeats weekly, so evidence accumulates fast. A mistake is cheap, because it happens in a draft and costs a paragraph. And access is narrow: read on the knowledge base and calendar, write on drafts, no write access to the CMS.

The side effect was the one described above. Trying to write down what "done" means for an article was the first time it became visible how much of that had only ever existed in my head. The thirteen rules are the result, and they would have been useful even without an operator.

What this means for you

Choose the first routine by provability, not by how much it hurts. The task that annoys you most belongs in round two or three, once the mechanism stands and you have had practice writing standards.

And if working through the list shows that none of your candidates cleanly meets the first criterion, that is not the end of the exercise. You have just found the actual answer, which is that your next step is not an operator but a defined standard for the work you would most like to hand over.

Frequently Asked Questions

How do I find the right first task for an AI operator?

List five to ten recurring tasks, then score each against four criteria: is "done" describable, does it repeat at least weekly, is a mistake cheap, and can access be cut narrowly? Give one point per criterion — whichever task scores highest is your first routine.

Why isn't the task that annoys me most the right one to automate first?

Because it usually annoys you precisely because it's unclear, keeps changing, and is full of exceptions — it's poorly defined. Poorly defined work is the hardest thing to hand to an AI operator, since there's no standard yet for recognizing when it's "done".

What kinds of work don't qualify for the first AI routine yet?

Four kinds: work with no definable standard, one-off projects, work with a wide blast radius, and actual decisions. Rocket Routine OS classifies decisions by impact into Root, Trunk, Branch and Leaf; anything at Root or Trunk level stays off-limits permanently, the other three are just postponed.

What was Rocket Routine's own first automated routine?

Weekly content production — not the most time-consuming task, but the best fit. "Done" is describable through thirteen edit patterns plus FTT, the work repeats weekly, a mistake only costs a paragraph in a draft, and access stays narrow: read on the knowledge base and calendar, write only on drafts.