9

Define done

Module 4 · Operations. After this lesson you can: write a definition of done before any handover and slice any task until each piece is checkable in minutes.

Consider an operator who starts delegating real work to an AI tool. The natural instinct is to hand things over the way they would with a person: "Clean up the project tracker." "Sort out the travel notes." "Handle the follow-ups." The tool always responds, always produces something, and what it produces is often impressive and frequently not what was intended. Ask it to tidy a tracker and it might rename categories, merge items it judges to be duplicates, and archive things it judges to be stale. Confident, sweeping, half right.

The instinct is to blame the tool. But the root cause is elsewhere. "Clean up" is not an instruction. It is a mood. A vague brief does not produce a vague result; it produces a confident result aimed at whatever assignment the mood seemed to point toward. The machine must produce something, so it aims at a guess. Every vague brief is an invitation to invent.

The practice that prevents this is small: write the definition of done before typing the request. One sentence describing the state of the world after the task is finished, concrete enough to check. Not "clean up the tracker" but "every item in the tracker has an owner and a date, anything finished is marked finished, and nothing is deleted." The first thing most operators notice when they try this is uncomfortable: for half their requests, they cannot write the sentence. They do not yet know what they want. The machine was never the problem.

The second thing to notice is about size. Some done sentences are checkable in two minutes: read the result, compare, finished. Others would take an afternoon to verify, which means they never really get verified, which means the task goes out unchecked. Those tasks are not too hard; they are too big. The fix is always the same: slice. Cut the task until each piece has its own done sentence and one main way it could go wrong, something verifiable over a coffee. "Plan the offsite" becomes "list three venues with prices for these dates," then "draft the invitation for the one selected," then "build the packing list." Each piece small, each piece checkable, each piece unglamorous. Boring is what reliable looks like up close.

Working well with AI tools is mostly about knowing, before starting, what finished looks like, and refusing to hand anything over until you do. The discipline costs one sentence. The lack of it costs an afternoon, and on the wrong task, considerably more.

Stop here. Think about the last thing you asked an AI to do.

Could you have written down, in advance, what done looked like?

Try this nowUnder 30 minutes
  1. Pick the next real task you were going to hand to your AI. Before you type the request, write the done sentence: the state of the world after, concrete enough to check.
  2. If you cannot write it, that is the finding. Sit with the task until you can. You have just caught a failure before paying for it.
  3. If checking it would take more than a few minutes, slice. Cut the task into pieces that each have their own done sentence and one main way to go wrong. Aim for pieces you could verify over a coffee.
  4. Hand over one slice and check it against your own sentence. Not against your memory of what you meant. Against what you wrote. Notice any gap: that gap is where invention happened.

Two questions before you go

Answer, then say whether you were sure or guessing. Being honest about which is the skill being trained.