Chapter 4: How to Have a Productive AI Conversation

September 2, 2026

This entry is in the series Get on board with Artificial Intelligence

⚡ FAST TRACK

The single-shot fallacy

Almost everyone starts out treating an AI tool like a vending machine. Insert request, receive output, evaluate output, decide whether the machine is any good.

That is not how competent users work, and the difference in results is enormous.

Competent users treat it as a conversation with a fast, tireless collaborator who has no ego and no memory of being corrected. They expect the first response to be roughly seventy percent right. They spend the next three exchanges closing the gap. The whole thing takes four minutes and produces something a single prompt would never have produced.

The unit of work is the conversation, not the prompt. This chapter is about running one well.

The shape of a good conversation

Nearly every productive AI exchange follows the same arc.

1. You describe the situation and what you want. More context than feels necessary. Most of the leverage begins with this context.

2. It produces something roughly right. Do not evaluate this as a final product. Evaluate it as a first draft from a capable stranger who does not know your world.

3. You steer. Usually two to four short corrections. This is the part beginners skip.

4. You stress-test. You ask it to find the weakness, to argue against itself, or to name its assumptions.

5. You extract the artifact. You ask for the final version in the form you need.

Steps 3 and 4 are what separate a useful session from a disappointing one, and they cost about ninety seconds.

The five steering moves

These five moves handle the overwhelming majority of situations where an answer is not what you wanted.

Move 1: Constrain

The most common problem with an AI response is that it is too long, too general, or in the wrong register. Say so, in as few words as possible.

Half that length. Drop the introduction and the summary. Just the middle. Plainer. Write it the way you’d say it out loud. No bullet points. Prose. Assume the reader is an expert. Stop explaining basics.

You do not need full sentences and you do not need to explain yourself.

Move 2: Redirect to the good part

When one section of an answer works, preserve it. Point to that section and ask the tool to rebuild the rest around the successful part.

The third option is the only one that fits. Give me five more in that direction and ignore the rest. Keep the structure, replace all the examples with ones from healthcare. That opening line is right. Rebuild the rest around it.

Beginners often abandon the conversation and start again. Experienced users mine the partial success, which is usually the more productive move.

Move 3: Ask for options instead of an answer

When you are not sure what you want, stop asking for the thing and start asking for a menu.

Give me three genuinely different approaches to this, not three versions of the same one. Label each with what it’s trading off.

The phrase genuinely different is doing real work there. Without it, you tend to get three near-identical variations, because the model is drawn toward the safest answer and produces it three times.

Move 4: Turn it against its own answer

Accuracy often improves when you ask the tool to challenge its own answer, yet few casual users do so.

What’s the weakest part of what you just told me? Argue against this as strongly as you honestly can. What are you assuming about my situation that I haven’t told you? If this advice failed, what would be the most likely reason?

The model has no ego invested in the previous answer, so it will criticize it freely and often accurately. Half the time the critique is more useful than the original.

Be aware of the limit here: this is a genuine improvement in quality, not a substitute for external verification. A model checking its own facts against its own memory will happily confirm an invented one. Chapter 9 gives you the method for the things that actually matter.

Move 5: Make it interview you

When a task is complicated and you do not know how to explain it well, hand the job back.

Before you answer, ask me up to six questions that would let you give a much better response. One at a time.

This is astonishingly effective. It surfaces the context you did not know was relevant, and the resulting answer is frequently better than anything you could have specified up front. Use it for anything consequential: a plan, a policy, a difficult message, a decision.

Context accumulates, and so does drift

Everything earlier in a conversation influences everything later. This is mostly an advantage: you build up shared context, and by the tenth exchange the tool understands your situation well enough that short instructions produce excellent results.

But it cuts both ways.

Errors persist. If a wrong assumption entered on message three and you did not catch it, message twenty is probably still built on it. When something feels subtly off, ask: Summarize what you currently understand about my situation and what I’m asking for. Wrong assumptions become visible immediately.

Long conversations drift. Very long sessions tend to wander, repeat, and lose the thread of the original instruction. Tools handle this better than they used to, but the effect is real.

Old topics contaminate new ones. If you spent thirty messages on a technical problem and then ask for help with a birthday message, the tone will bleed.

When to start a new conversation

Start fresh when the topic changes substantially, when the conversation has become long and circular, or when you have corrected the same mistake more than twice without it sticking.

Carry your context forward deliberately. Ask the old conversation:

Write a paragraph I can paste into a new conversation that gives all the necessary background.

Then paste it into the new one. Thirty seconds, and you keep everything valuable and drop everything that was going wrong.

Three failure patterns and their fixes

The loop. You correct something, it apologizes, it makes the same mistake. Correcting again does not help — the failed attempt is now in the context, making repetition more likely. Fix: start a new conversation, and state the requirement as a positive rule up front rather than as a correction. “Write this in plain language at a sixth-grade reading level” works better than “stop using jargon.”

The flatterer. Every idea you propose is excellent, insightful, and a great question. This is a known tendency in these systems and it is dangerous when you are using AI to evaluate your own thinking. Fix: ask adversarially. Not what do you think of my plan but here’s a plan, find its three biggest weaknesses. Better still, present it as someone else’s: A colleague proposed this. What would you push back on?

The plausible invention. The answer contains a specific citation, statistic, quotation, or legal provision that does not exist. It will be perfectly formatted and entirely fictional. Fix: this is not fixable through conversation. Chapter 8 explains the mechanism and Chapter 9 gives you the verification method. For now: any specific fact you intend to rely on gets checked outside the tool. Every time.

Ending the conversation well

Close deliberately by gathering the useful work from the conversation into one clean result:

Give me the final version, clean, with nothing else around it.

Or, when the thinking mattered more than the output:

Summarize what we worked out, the decisions I made, and what’s still open.

Or, when you will do this again:

Write me a reusable prompt that would get this same result next time, with blanks where the details change.

That last one quietly builds your personal prompt library over a few months, which can eventually become part of your personal system.

Exercise 4.1: Run the interview move

Pick something real and consequential: a plan, a decision, a difficult message you need to send.

Instead of describing it fully, open with: Before you answer, ask me up to six questions that would let you give a much better response, one at a time.

Answer honestly. Compare the final result to what you would have gotten from a single request.

Exercise 4.2: Turn it against itself

Take any answer an AI tool has given you this week and ask all four of the Move 4 questions.

Write down anything the critique surfaced that you had not considered. For most people this exercise is the moment the tool stops being a novelty and starts being useful.

Exercise 4.3: Catch the drift

Find a conversation of yours that ran more than fifteen messages. Ask: Summarize what you currently understand about my situation and what I’m asking for.