The tasks where the model is the wrong tool, the ones a free rule does better and faster, and the four questions that decide between them. This chapter turns ten chapters of instinct into a card that fits above a desk — and ends with the one thing that has to stay human for any of it to work.
Nine chapters have been about getting more out of the model. This one is about the tasks where the right move is not to use it, and it exists because those two skills are the same skill. A recommendation is only worth having from someone whose recommendations are sometimes no.
Four pages, and they build one artifact. First the weak ground, mapped as three patterns rather than a list of exceptions — exactness, facts that must be true, and rules that must fire identically — all three falling out of a single sentence from Chapter 1. Then the cost question a capable user starts asking once the novelty wears off: not whether the model can do it, but what the cheapest thing is that does it well enough, with a ladder of four rungs and Tessa auditing her own bill on the way past. Then the card itself, four questions in a fixed order, each one derived from the chapter that taught it — this course folded to the size of a postcard. And last, the quiet warning: every habit in this book assumes a human who can still tell good from bad, and that is a skill which has to be kept on purpose now that nothing keeps it by accident.
Which tool the task is actually asking for
Reach for the model
Drafting, reshaping, sorting a pile whose sorting needs an actual reading, explaining a clause in plain words.
Reach for something else
Exact arithmetic, facts that must be true, and rules that must fire the same way every single time.