A diagnostic lens is a reusable question for finding why a prompt fails, varies between runs, or produces plausible but unsuitable output. Check the prompt before assuming the cause is the model or an external factor. section 1.6 section 1.6.3
Use three questions to inspect a failing prompt:
- What contract did I specify? section 1.6
- What context did I omit? section 1.6
- Where did I leave too much room for variability? section 1.6 section 1.6.3
When results vary, look for underspecification: multiple interpretations may fit the prompt. An output schema, constraint, or worked example can narrow that interpretive space. Some variance may instead come from sampling, so distinguish prompt issues from model-level controls. section 1.6.3
Support-ticket classification
If tickets using feature names classify reliably but user descriptions do not, add a reference mapping common descriptions to the internal product taxonomy. The missing context, rather than model malfunction, explains the inconsistent results. section 1.6.3
