Article

3 Oct 2026

The questions your AI should be asking you

The fastest route to better AI output is letting it ask the questions first. One simple prompt turns a one-line request into a draft you'll keep.

Cartoon: a robot in glasses interviews a sweating manager who is holding a single sticky note

Here's a scene I watch play out in almost every training room, usually within the first hour. Someone has a real job to do, a reference letter for a former colleague, say, or a summary of a messy project for the board, and they type a single line into the AI tool, hit enter and lean back with the hopeful expression of a person buying a lottery ticket. What comes back is fluent, tidy and completely interchangeable with anything else on the internet. It could describe anyone. They read it, sigh, and offer some version of "well, it's not quite there, is it?", and underneath that polite verdict is a more personal worry: everyone else seems to get brilliant results from these tools, and somehow they've missed the trick.

The trick they've missed is tiny, and anyone can learn it before lunch.

Why the first draft so often disappoints

What I've found, working with teams through their first few months with AI, is that the gap between a mediocre draft and a good one almost always comes down to what the person knew and didn't type, the context that lives in their head and feels too obvious to mention. You know why this reference matters. You know what the person was brilliant at, where they struggled, and how you feel about them leaving. The AI knows none of it, so it fills every gap with the safest, most average guess available, and you get the most average letter available.

Most of us would struggle to write that context down unprompted, if we're being honest, because we don't know which parts matter until someone asks. It's the same reason a decent GP lets you finish describing your symptoms before reaching for the prescription pad.

So let the AI do the asking.

Turning the conversation round

The prompt I use for almost any substantial piece of work looks like this:

I am trying to achieve [your goal]. I want you to act as my expert [relevant role] and help me achieve it. Before we solve it, I need you to understand the problem. Ask me [number] questions, all at once, to gather as much context as you can. After I've replied with the answers, [the action you want].

For the reference letter, that becomes: "I'm trying to write a recommendation letter for a former employee. Act as my expert HR director. Ask me five questions, all at once, and once I've answered them, write the letter."

The role gives the AI a point of view, so its questions come from someone who has written a hundred of these letters. The number keeps things proportionate: five for a letter, ten or more for a strategy paper. And "all at once" spares you a drip of single questions, which feels a lot like a particularly slow online form interviewing you.

What came back were the questions an experienced HR director would ask. Who is the person and what did they do? Who is the letter for? What are their strongest qualities? How do you want it to sound? And what's your candid view on why they left?

Answering took under five minutes. Writing the letter from scratch would have taken me half an hour, and the draft that came back beat my own first attempt comfortably, because every line of it grew out of my answers.

The question you'd never have volunteered

Nobody types "here's how I really feel about them leaving" into a prompt, and yet that feeling shapes everything about a reference: how warm it is, what it leaves out, which strengths it leads with. You'd have written the whole letter around that feeling anyway, just without telling the AI, and then wondered why the tone came out slightly wrong.

Ask an AI to help with a price increase letter to a long-standing client, a tricky supplier email or a board paper, and the most valuable question it asks will usually be the one that makes you pause before answering. In that pause, your judgement enters the work, and judgement stays firmly with you.

After a few weeks of answering these questions, you start to anticipate them, and your opening prompts get richer because you've absorbed what a good brief needs. The framework teaches you to brief well in the first place, which makes it one of the few tools I know that gets less necessary the more you use it.

Knowing when to pick up the pen

The letter wasn't perfect. Somewhere in the middle, Claude produced the line "the numbers matter, but the story behind them matters more", the sort of sentence that sounds profound on first reading and means nothing on the second. Classic AI filler. I deleted it and wrote my own line, which took about ten seconds.

Most of us, faced with one bad line in an otherwise good draft, reach for another prompt out of habit, and five minutes later we're on the fourth rewrite of a sentence we could have fixed in ten seconds. A strong first draft exists to free you up for the parts only you can do, the judgement calls, the nuance, the relationship behind the letter, and knowing when to take the pen back is part of that.

I ran the whole exercise on Claude Sonnet at medium effort, which handled it comfortably. The bigger models come into their own on longer, more complex jobs, and a reference letter falls well within the everyday model's reach.

Where to start

Pick one piece of writing on your list this week that matters enough to get right, the proposal, the awkward email, the update you keep putting off, and run it through the prompt before you write a word. Answer the questions properly, edit the draft by hand where it needs it, and notice which question made you stop and think.

Five questions and five minutes produced a letter that sounded like it came from someone who knew the person. What strikes me, every time I run this with a team, is how much people discover they already knew once someone asked them properly. Which leaves an uncomfortable thought for any leader: if five questions from a piece of software can draw that much out of you about one reference letter, how much does the team know about your business that nobody has ever thought to ask?

Where to go next

Try the framework on one of your own prompts with the free Prompt Pit Stop. AI-101 Training teaches it to your whole team, live, with their real work, and How to get more out of Claude puts it alongside skills and connectors.