Process19 Aug 2026

Split AI work into asking, doing and checking

By Alan Goh · 2 min read

Three roles

Ask, do and check, separately

  • Decide, and write down what good looks like
  • A fresh session does the work
  • Someone else checks it against your description
Short answer

AI work goes better with three roles kept apart: you decide and write down what good looks like, a fresh AI session does the work from that instruction alone, and someone with no stake in the result checks it against your description.

Think about how a renovation runs. You decide what you want. A contractor builds it. Before the final payment, an inspector walks through and checks the work: someone who didn’t build it and has no reason to defend it.

Three roles, each kept separate. Most AI work quietly folds them into one.

One conversation, three jobs

A typical session goes like this. You open a chat, think out loud, change your mind twice, ask AI to produce the thing, and then ask it whether the thing is good.

It feels natural, and it puts three jobs in the same place: deciding what’s wanted, doing the work, and checking it. The check carries all the assumptions of the conversation that produced the work. So does the person reading it, who is already hoping it’s right.

Give each role its own space

You decide. Think it through, then write down two things: the instruction, complete enough to stand on its own, and a short description of what a good result looks like.

A fresh session does the work. Start a new AI conversation and give it only the written instruction. With none of the earlier back-and-forth, it works from what you decided rather than everything you considered.

Someone else checks. A colleague, or another fresh session, compares the result with your description of good. They have no stake in the work, so they simply report what they find.

The inspector inspects, and the contractor fixes.

The written instruction does most of the work

Splitting the roles means the instruction has to be written down, whole. That turns out to be the most valuable part.

It also gives you a quick test. Run the instruction in a fresh session with no background. Whatever comes back wrong points to something the instruction left unsaid.

And write what good looks like as well as the steps. Steps let a checker confirm the instruction was followed. A description of good lets them say whether the result is fit to use.

Try it once

You can start with one piece of work. Next time something matters, such as a proposal, a report for the board or a policy draft, write the instruction and the description of good, run it in a fresh session, and ask someone else to check the result against your description.

Then compare what the check finds with what you’d have caught on your own.

In your department

  • Strategy & Planning: For a board paper or a proposal, write down what a good result looks like before you start, and ask someone outside the work to check against it.
  • Accounts & Finance: Finance already keeps preparing and approving apart. Give AI work the same split: one written instruction, one fresh run, one independent check.
  • Operations & Delivery: Give a recurring AI task a written instruction and a named checker, the way a job gets a supervisor’s sign-off.

Want to try this with your team?

Tell me what your team is working on, and I'll suggest where to start. I read and reply to every message myself.

More notes

  • 05 Aug 2026

    AI can be confidently wrong

    AI writes by predicting what fits, so it can give an answer that is specific, fluent and wrong. Treat every AI answer as a draft, and check the parts that carry weight, such as numbers, references and names, before you rely on it.

    Read it →
  • 15 Jul 2026

    Brief AI the way you'd brief a person

    Search engines taught us to type a few keywords. AI works from meaning, so it gives better answers when you describe the situation and what the result is for, in plain sentences, the way you would brief a person.

    Read it →

All the notes →