Not a list of prompts. A way of deciding, task by task, whether the machine belongs in the loop at all.
Most advice about using AI at work starts with a product and looks for something to do with it. That order is backwards, and it is why so many pilots quietly die after a month. Start instead from a task you already do, and ask three questions about it.
Can you tell, quickly, whether the output is right? Drafting a first version of a meeting summary passes: you were in the meeting, so a wrong summary is obvious in seconds. Producing a client's billing figure fails: checking it costs as much as doing it, so nothing is saved.
Does being wrong cost more than being slow? Where an error is cheap and reversible, a fast imperfect draft is a real gain. Where an error is expensive or hard to notice later, the time you appear to save gets borrowed against a bill you pay later.
Is the work yours to hand over? Confidential material, personal data about other people and anything under a client agreement are not yours to paste into a third party service just because it would be convenient.
The consistent wins are unglamorous: turning rough notes into a structured first draft, rephrasing something you already understand for a different audience, generating the boring variants of a text you will then pick from, explaining an unfamiliar piece of code or a dense document well enough for you to ask better questions about it.
The consistent losses are also stable: anything requiring a fact the model has no reliable access to, anything where the tone of authority hides the absence of knowledge, and anything you cannot check faster than you could have done it.
The framework above is a way of deciding. This is one measurement behind it, run on this site on 2026-08-10: three low cost models extracting seven fields from 24 supplier invoices, at four levels of degraded text. The number that matters is not accuracy. It is how often a wrong answer came back with no warning of any kind, in a task where a human check was supposed to be the safety net.
| Degraded text | Wrong extractions out of 72 | Wrong with no warning | Share silent |
|---|---|---|---|
| N1 clean | 0 | 0 | not applicable |
| N2 columns lost | 0 | 0 | not applicable |
| N3 totals shuffled | 21 | 3 | 14.3 percent |
| N4 labels lost | 21 | 10 | 47.6 percent |
Read it against the first question of the framework. On clean text the models were perfect, 144 extractions out of 144, and the check would have found nothing to do. On degraded text they broke, and the silence grew with the difficulty: the harder the document, the less the model said about its own doubt. That is the exact shape of the failure the three questions above are meant to catch, and it is why "can you tell quickly whether the output is right" is the first of them and not the third.
See the full dataset, with the raw responses, the ground truth and the scripts.
The articles in this section work through that decision framework and its failure modes in detail.
Ordinary software breaks loudly. A language model does something stranger: it produces a confident, well formed answer whether or not it has one. Here is where that bites, and what catches it.
Read the articleMost advice starts with a tool and looks for a use. Start with the task instead, and put it through three questions. Two of them are about checking, and that is the point.
Read the articleTaken from what shows up alongside this search, answered short.
One email when a new article is published. No tool of the week, no affiliate list, no forwarding of your address to anyone.
