Lesson 7 of 8

When a prompt fails

6 min read

The answer comes back wrong and the reflex is to hit send again, hoping for a better roll. But bad output almost always has a reason — and once you can name how it failed, the fix is usually a single line in your prompt, not luck.

Read the failure, don't just re-roll

Start by naming how it went wrong. It made a fact up. It ignored a bit of the ask. It drifted off-topic, or handed you a wall of prose when you wanted bullets. The most common — and most dangerous — is the confident [hallucination](glossary://hallucination): a made-up detail delivered in the same sure voice as the real ones. Spotting which failure you're looking at is the whole first step.

Bad output comes in recognisable shapes: made-up facts, ignored instructions, wrong format, off-topic drift. Naming the failure is what points you at the fix.

Fix one thing, then re-run

Each failure maps back to a prompt cause: vague output means too broad an ask; an invented fact means you demanded something it couldn't know; the wrong shape means you never named a format. A sneaky one is an unclear "it" or "that" — the AI can't tell what you meant. So change the one thing that matches the failure and re-run. And if wording tweaks stop helping, the problem may be the technique itself.

Match the failure to its cause, change that one thing, and re-run. When small edits stop working, the fix is a different technique — not louder wording.

If tightening the wording keeps not working, stop tightening. Jumping from zero-shot to a couple of examples, or adding "think step by step," fixes problems that no amount of rephrasing will.