theHermeticum
PLES Subscribe

Exercise - five minutes

Five questions
before you ask AI

TL;DR

Before you hand a problem to a model, answer five questions on paper: what you are actually asking, what you know and what you assume, your own answer, what you will use the reply for, and what you will not delegate. It takes five minutes and gives you a position of your own, so the reply becomes something to weigh rather than a verdict. After the exchange, one line: did the answer change my position, or only my confidence?

The moment to keep your judgement is before the model answers, not after. Once a confident reply is on the screen, it becomes the thing you react to.

Five questions fix that in five minutes. Write short answers, on paper or in a note. The point is not a good document. It is having your own position before another voice arrives.

The five questions

  1. What are you actually asking? Write the question in one sentence, without jargon. If it takes three sentences, you are asking three questions; pick the one that matters today.
  2. What do you know, and what are you assuming? Two columns: facts you could show someone, and assumptions you are treating as facts. Most bad delegation starts here.
  3. What is your own answer, before you ask? Even a rough one. It is the position the model will argue with, and the only way you will notice if the reply quietly replaces your thinking.
  4. What will you use the answer for? A decision, a draft, a check, a second opinion? Naming the use decides how much weight the answer deserves.
  5. What are you unwilling to delegate? Judgement, taste, the relationship, the final call, the responsibility. Write it down before you start, not after.

After the exchange, one line: did the answer change my position, or only my confidence?

A worked example

Hypothetical. It follows the scene that opens the technology chapter of Modern Hermet, which the book itself calls a composite: you have two offers for a client and want a model to tell you which one to send.

  1. The question. Which of my two offers goes to this client on Friday?
  2. Known and assumed. Known: the client asked for a quick start; offer A is cheaper and slower, offer B is faster and dearer. Assumed: that price matters more to them than time, and that they will not negotiate.
  3. My answer first. B, because they asked for speed, though I am not sure the budget stretches that far.
  4. The use. A second opinion on a decision that goes out under my name, not the decision itself.
  5. Not delegated. The final choice, and the message to the client.

Only now ask, and ask for input to your judgement rather than a verdict: the arguments against B, the objections the client might raise, what you have not considered. The chapter’s rule: “Delegate the input, never the output.” Then write the line. Here the reply raised payment terms, which you had not thought about; you still send B, with a different payment schedule. The position moved a little and the confidence did not jump, which is what the exercise is for.

Why it helps, and where it stops

An answer written first gives you something to check the reply against, and naming the use sets how much weight the reply gets. The technology chapter points to research on automation bias, Parasuraman and Riley (1997) and Skitka and colleagues (1999): people working with automated prompts miss what the system did not flag, and carry out what it flagged wrongly. The five questions are our response to that pattern. No study cited here tests this exercise, and it does not replace knowing the field: if you cannot spot an error in the answer, no written question will spot it for you.

Take it with you

The five questions fit on one page: download them, no sign-up. The wider argument, what to hand to AI and what to keep, is in human agency in the AI age and in the open chapter, Technology, free in full.

Fair questions

Do I need to use AI for this?

No. The questions work before any decision you are about to hand over, to a model, a colleague or a search engine. With AI they matter most, because its replies arrive sounding sure.

How long does it take?

About five minutes. One line per question is enough; the point is a position of your own, not a document.

Has this exercise been tested?

Not as such. It rests on a documented pattern, automation bias, and on an ordinary habit: writing your own view down before you hear another. No study cited here tests these five questions.

Sources & further reading

  • Modern Hermet (The Hermeticum, 2026), chapter 8, Technology (open in full): the scene of the two offers and the rule quoted on this page.
  • Parasuraman and Riley (1997) and Skitka and colleagues (1999) on automation bias, as cited in chapter 8.
  • The five questions are our own arrangement: Pre-AI Orientation (The Hermeticum, 2026). They are not a treatment and make no therapeutic claim.