Ask AI What It Assumed

No prompt is ever complete — and unlike a colleague, AI almost never asks. It quietly fills the gaps: it assumes who the text is for, how formal it should be, that "the team" means five people, and that "the budget" means the annual one. The answer then looks confident and usable, except it's standing on foundations you never approved. The cheapest quality-control habit there is with AI is one follow-up question after every answer: "what did you assume on your own?"

How to do it

1. Let the model answer your task normally — a plan, a proposal, a calculation, an email.
2. Then send this question:

Before I use your answer: list the assumptions you filled
in about the request on your own, because I didn't state
them explicitly.

For each one, note:
- exactly how it shaped the answer,
- what would change if the opposite were true.

Rank them from most reversible-and-damaging — i.e. starting
with the assumption whose change would break the answer
the most.

3. Go through the list and correct the wrong assumptions in one sentence: "point 2 doesn't hold, the audience is technical staff, adjust the answer."
4. For important tasks, make the question part of the original prompt: "always list your assumptions at the end of the answer." The model then plays with its cards face up from the start.

A typical scenario

Manager Radek has AI draft a rollout plan for a new time-tracking system. The plan looks great: timeline, communications, training. He'd send it onward — but first he asks the assumptions question.

Three things fall out of the list that weren't in the original prompt: the model assumed everyone works in the office (half the staff is in the field without a computer), that internal IT handles the rollout (a vendor actually does), and that it's a single location. The first assumption is the most damaging — the entire communications plan rests on emails that field staff never read. Radek fixes two sentences in the prompt, and the second version of the plan is usable. Had he sent the first version onward, it would have come back a week later from operations, with comments attached.

What you get out of it

You catch bad foundations in a minute instead of a week — and you often learn something about your own prompting too, since the list of assumptions is a precise imprint of what was missing from the request. The question costs nothing and works on any task, from an email to a strategy. It's a different discipline from fact-checking (there's a separate guide for that): here it's not about whether the model is telling the truth, but whether it's solving the same problem you are.

Pro tip

The question is most valuable for anything involving numbers: calculations, estimates, comparisons. Explicitly ask for "the assumptions behind the calculation — exchange rate, rates, period, what's included and excluded from the price." That's exactly where the quietest assumptions hide: the model might happily calculate with tax you never mentioned, or last year's exchange rate — and without the assumptions spelled out, you'd never spot it in the final number.