AI Fluency.

The framework

1. 01

   It's weakest exactly where your work is most specific.

   Reliability tracks how much text exists about a thing, and that has almost nothing to do with how hard the thing is. Widely written-about material — a standard contract clause, a well-documented framework, a language half the internet codes in — comes back solid. Your sector's pricing quirks, your own numbers, a rule that changed last month, anything sitting behind a login: thin to the point of invention. You can run that check before you ask rather than after. How much has been written in public, at length, about exactly this? If the answer is not much, you aren't asking a hard question. You're asking one there was never any material for.

2. 02

   Treat every session as an amnesiac's first day.

   There's no continuity between conversations, and inside one there's nothing beyond the text in front of it. The memory features in consumer products are a filing system bolted on the outside: notes stored and pasted back in, not the model recalling you. So a brief is an asset you keep, not a speech you improvise. Store your standing context in one file — the business, the reader, the constraints, the standard of done — and paste it at the top. The operator who says the model keeps forgetting is usually retyping that context badly, differently, every time.

3. 03

   What happens when you send the identical prompt twice?

   Nothing in the machinery guarantees the same answer twice. The model works out a distribution over what could come next and samples from it, so two runs of the same words can diverge — and where they diverge is information you get for free. Send anything load-bearing twice, in separate sessions, using the same words rather than two framings of the question. Agreement on a figure is no proof it's right, but disagreement is proof that something is unsettled, and it points straight at the sentence to check first. The property that makes the machine unreliable is the one that shows you where it is unreliable.

4. 04

   Check that the source exists before you check whether it's right.

   A search engine that fails hands you the wrong real document. A model that fails hands you a document-shaped string: a journal that exists, an author who publishes in the field, a year, a page number, and no such paper. That is a different failure and it takes a different first move. Follow the link, search the title inside quotation marks, look for the page number. Whether the source says what the answer claimed is the second question; whether it exists at all is the first, and only generation produces that failure.