Lesson 11: Reading and diagnosing a bad output
When the reply disappoints, the first mistake is to blame the model. The second is to rewrite the prompt at random. The right move: diagnose. Name the failure — too long, wrong format, an invented fact, an ignored constraint — and trace it back to the prompt: what was missing there? In this lesson w
A bad output isn't a dead end — it's a diagnosis. If you got a long paragraph instead of 3 bullets, the output is telling you what the prompt forgot: format and count. Name the failure, add the missing piece to the prompt, and run again.
- diagnosing output
- Reading a disappointing reply to name the failure and map it back to what the prompt failed to ask for.
- failure mode
- A recurring kind of bad output — too long, wrong format, an invented fact, an ignored constraint, or answering the wrong question.
- hallucination
- When the model invents a fact or figure that sounds plausible but isn't true, often when the prompt gave no source or context.
