Stop asking it to write the thing, start asking it to take the thing apart

Most people use these systems like a vending machine. Type the request, take what comes out, get on with the day. The output is broadly acceptable, occasionally good, and always slightly generic, and after a few months the conclusion settles in that this technology is fine but a bit overrated.

There is a control sitting one layer up from where they are typing, and it changes the character of the thing entirely. It is not a secret and it is not technical. It is the instruction that establishes who the system is being before it reads a single word of your actual question.

Compare "analyse this data" with "approach this dataset as a forensic accountant would, looking for patterns that suggest something has been smoothed over, and stay sceptical of anything that looks unusually tidy". Same underlying system, same data, and the second one is a different colleague. Not because the words are magic, but because you have told it what to be suspicious of, and suspicion is the thing generic prompting never supplies.

The part I want to argue about#

Here is where I diverge from most of what gets written on this subject, including things I have written myself.

You should not be asking these systems to produce your work. Not because it cannot, but because of what happens next. When you stand in front of a board and someone asks why you recommended a particular approach, you cannot say "hold on while I ask why it suggested that". You own every word, every assumption, every number. The authorship is not a formality, it is the thing that makes you accountable, and accountability is most of what you are actually being paid for.

What you can do, and almost nobody does systematically, is build the critic you never get access to.

Think about the last proposal you took to senior management. You spent hours on it, anticipated objections, marshalled evidence. How much of that did you actually test against a hard-nosed commercial mind before you were standing in front of one? Almost none, because the person who would give you that scrutiny was busy, or it was Thursday, or asking felt like admitting you were not sure.

That is the gap. Not writing, marking.

What makes a critic useful rather than decorative#

The failure mode here is generic role description, and it is why most attempts at this feel like a party trick that wears off.

"Act like a CFO" gets you an encyclopaedia entry with opinions. It knows what CFOs care about in the abstract and will produce plausible CFO-adjacent commentary that improves nothing, because the observations are the ones you already anticipated.

What makes the difference is specificity that has teeth. Not a job title but a temperament: someone who has watched too many businesses fail while profitable on paper, who becomes visibly impatient with any sentence containing "once we scale", who asks what happens if you achieve thirty per cent of that projection rather than a hundred, who wants best case, worst case and most likely before looking at the headline figure at all. Give it priorities in order, because a critic without priorities just lists everything. Give it a manner, because how criticism arrives affects whether you act on it. Give it boundaries, so it is not obstructive for the sake of it, and so you can tell the difference between a genuine objection and reflexive negativity.

The test is whether it says something you had not thought of. If your critic only ever confirms your own worries back at you, you have built a mirror.

The three that earn their keep#

A handful of these do the vast majority of the useful work, and I would rather have three sharp ones than fifteen vague ones.

The commercial sceptic, who attacks the money. Where does cash come from in month three, what is the customer acquisition cost you have not mentioned, which of these figures is an estimate wearing the clothes of a fact.

The regulatory and risk mind, whose job is to notice what you have accidentally promised. Words like guarantee, always and never are load-bearing in ways people do not intend, and someone whose instinct is to ask "could we actually be held to this" catches problems while they are still free to fix.

And the strategic alignment check, which is the one most people skip and which quietly kills the most proposals. Organisations have stated priorities. Proposals fail not because they are bad but because nobody demonstrated how they serve those priorities, and the person reading it should not have to do that work on your behalf. A critic configured with your actual organisational objectives, evaluating whether your case connects to them, is unglamorous and remarkably effective.

The pattern worth noticing#

There is a secondary benefit that takes a couple of months to appear and is worth more than any individual review.

If you keep the feedback, patterns emerge. If the commercial critic challenges your timelines every single time, that is not a quirk of the tool. That is a systematic tendency in your thinking, and now you know about it. If the risk perspective repeatedly finds unsubstantiated claims in your work, you have learned something about your own drafting habits that no annual review was ever going to tell you.

I have found this genuinely uncomfortable, in the way useful things often are. It turns out I am consistently optimistic about how quickly people will adopt something, which I half knew and had never seen laid out across a dozen documents.

Keep the authorship#

This whole approach rests on one distinction, and it collapses if you lose it: the critic prepares work for your judgement, it does not replace your judgement. It spots the weak assumption. You decide whether to strengthen the argument, change the plan, or proceed anyway because you know something the critic does not.

That last option matters more than it sounds. Sometimes you are right and the objection is wrong, and being able to say why is exactly the confidence you wanted from the exercise.

Homework marker, not homework writer. Everything useful about this follows from getting that the right way round.