AI Fluency.

The framework

1. 01

   Make 'I don't know' a legitimate answer, in writing.

   A model won't volunteer uncertainty into a request that has only asked for an answer. So put the permission in the brief in plain terms: if the material doesn't settle this, say so and stop; if you're guessing, label the guess. It reads like a small courtesy and it changes the output, because you've altered what counts as success for that request. The version that fails is the vague plea to be accurate. The version that works names abstention as an expected, unpenalised outcome and asks for it in a specific place in the answer.

2. 02

   Confidence claim by claim, not one rating at the end.

   A single rating on a long answer is close to useless, because the solid parts and the invented parts are mixed together and average out to something reassuring. Ask instead for the answer marked up line by line — high, medium, low, with a reason for anything below high — and read the low ones first. Stated confidence is itself generated text and not a measurement, so never treat it as a probability. Treat it as triage. It is often good enough to tell you where to look first, which is all you needed from it.

3. 03

   Separate what you can check from what you can't, before you send.

   Split the request in two. Which parts of the answer could you check inside ten minutes — a figure in your own system, a clause in a contract you hold, a name you can search? Which parts would you have to take on trust? The second list is where fabrication does its damage, and the right move is usually to take those claims out of the request rather than to ask more politely. A claim you cannot check and cannot remove isn't a task for a generative system at all; it's a phone call to somebody who knows.

4. 04

   The tell everyone waits for doesn't exist here.

   Hedging, hesitancy, a wobble in the prose — the signals you'd read off a nervous colleague — are produced independently of whether the content is right. A marker you asked for at least sorts the answer into an order you can work through; unprompted throat-clearing sorts nothing at all. Either way the calibration has to come from outside the answer. Keep a short log of every occasion a model was wrong on your work: what kind of question it was, what it invented, how you caught it. A dozen entries in, you'll hold something no amount of reading about the technology gives you, which is a specific account of the ways it fails on your material.