Ask How First

A model will do what you asked. That sounds like the good news, and for the most part it is, but it contains a trap that gets more expensive as the model gets more capable: it will also do what you asked when what you asked for was wrong.

It does not stop to wonder whether you have thought this through. It has no way to know that the thing you described is the third-best route to the thing you actually want, and no inclination to ask, because a request that can be fulfilled is a request it will fulfill.

So put one step in front of the work.

Don't build anything yet. How would you approach this, and what are you assuming?

What comes back is cheap to reject

The first argument for this is arithmetic. A plan is a paragraph. The work is not.

Reading a paragraph and saying no, not like that costs you thirty seconds. Reading four hundred lines, or a finished report, or a restructured spreadsheet, and discovering that the premise underneath it was wrong costs you the reading, the unpicking, and the sunk feeling that makes people keep bad work rather than throw it away. That last cost is the real one, and it is why so much wrong work survives.

Rejection has to be cheap or it does not happen. Asking for the plan is how you make it cheap.

The assumptions were never going to be mentioned

The second argument is the one that catches people out, because the failure is silent.

Every substantial request has holes in it, and a model fills holes the way it fills everything else — with the likeliest thing. It assumed the data was clean. It assumed you meant all customers rather than the ones who have bought twice. It assumed you had a budget, or that you did not. It assumed the deadline was soft.

None of that gets said. It is not being cagey; there is simply no mechanism that flags an assumption as an assumption. The gap gets filled and the work proceeds on top of the filling.

Ask for the approach and the assumptions surface as text, where you can look at them. Half of what you will find is fine. The other half is the reason the finished thing would have been useless, arriving early enough to be a sentence rather than a rebuild.

Sometimes it knows a better way

The third argument is the one people do not expect, and it is worth the question on its own.

You describe a route because a route is what you have. You have been thinking about your problem for a week and you have arrived at a plan, and the plan is a reasonable plan for someone who knows what you know. What you asked for is that plan, executed.

Ask how would you do this and you sometimes get: the usual approach is the one you described, but given the constraint you mentioned in passing, there is a shorter path. That answer was always available. It simply never had an opportunity to be given, because you asked for your plan and your plan is what it was going to build.

This is the difference between an instruction and a consultation, and the only thing standing between them is a question you did not ask.

The plan becomes the specification

Once you have corrected the approach — one sentence, usually, and often none — you have something more useful than a corrected approach. You have a written agreement about what is being built, produced before anything was built, in language you both just read.

That artifact is worth keeping. When the work comes back you have something to check it against that is not your memory of what you wanted a week ago. When it drifts, and long work drifts, you have the thing it drifted from. When you hand the job to someone else, or to a fresh conversation with no memory of this one, the plan is what you hand over.

None of which requires any theory about how models work. It is the oldest professional habit there is, wearing new clothes: tell me how you plan to do it before you start doing it. We ask this of contractors, of surgeons, of anyone whose work is expensive to undo. The only reason it feels novel here is that the work looks free, and it is not — it is just cheap enough that we stopped asking.