Diagnosing Bad Outputs

When an output is wrong, the instinct is to rewrite the prompt from scratch. That instinct is expensive and slow. Before rewriting anything, diagnose. Every bad output fails in a specific way, and each failure mode has a specific fix. Rewriting without diagnosis means you might fix the wrong thing — and introduce new gaps while doing it.

1. Wrong format The output is correct but in the wrong structure — prose when you needed bullets, a list when you needed a table. Fix: add explicit output format specification. This is the easiest failure to diagnose and the easiest to fix.
2. Wrong scope The output covers too much or too little. Too much: the model determined thorough meant comprehensive. Too little: the model interpreted the task narrowly. Fix: add explicit scope constraints — "address only X," "cover A, B, and C," "limit to 200 words."
3. Wrong tone The output is too formal, too casual, too technical, or has the wrong voice. Fix: add measurable tone specification — reading level, formality level, person (first/third), use of jargon. Vague tone instructions produce vague tone compliance.
4. Wrong depth The output is either surface-level when you needed depth, or detailed when you needed a summary. Fix: add depth instructions — "provide executive-level overview," "go into technical implementation detail," "include supporting evidence for each claim."
5. Hallucination The output contains invented facts, fabricated citations, or statements that cannot be verified from the provided source material. Fix: add grounding constraints — "use only the information provided in the document below," "if you do not know, say so," "do not extrapolate beyond the source." Hallucination is a constraint failure, not a model failure.