Getting Real Work out of a Model · Chapter 06
Four reasons an answer is bad
1:55 · Narrated with a synthetic voice. Captions are available in the player.
Read the transcript
When an answer is bad, name the cause before you fix it. There are four. It did not have what it needed, and the symptom is an answer true of a category rather than of your case. It misread what you wanted, and you got a good answer to a different question. The output is in an unusable form, where the content is right and the shape is wrong. Or the task is beyond the tool. Work through them in that order, because the first is the commonest and the cheapest to fix. And learn the tell for the fourth: better prompting produces a better-looking wrong answer. If your third attempt is more polished and equally wrong, stop rewriting and change the approach. For a prompt you will reuse, change one thing at a time against the same fixed items, several runs each, and then keep or discard explicitly. The discard step is the one everybody skips, and it is why inherited prompts grow to twelve elements nobody dares remove. Keep five to ten items forever. Two ordinary ones, three or four hard ones taken from real failures, and one whose correct answer is a refusal. That last matters because the ability to correctly not answer disappears silently, and its failure produces a fuller-looking result rather than a broken one. Run them after any change to the prompt, after a model update, and monthly if the prompt runs unattended.