Prompt engineering

Prompt engineering

A prompt is a specification. Vague specs get confident garbage; the fix is almost never "more words", it is the right words in the right structure.

Method

1. Write the success criteria first. What does a correct output contain, in what shape, and what would make it wrong? If you cannot check an output, you cannot prompt for it.
2. Pick the technique for the task type:
   • Classification / extraction → 2–5 worked examples (few-shot) showing input → exact expected output, including one tricky case. Examples teach format and edge handling better than any description.
   • Reasoning / math / multi-step → instruct step-by-step thinking before the answer, and separate the reasoning from the final answer so it can be parsed.
   • Creative / stylistic → role and audience ("you are a…, writing for…"), two or three constraints that define the voice, and one example of the tone if you have it. Constraints beat adjectives.
   • Structured output → show the exact schema with a filled example. State what to do when a field is unknown (empty string? null? omit?) or the model will invent.
3. Structure the prompt in blocks, clearly delimited: role/context → task → rules → examples → the input (fenced or tagged, e.g. <document>…</document>) → output format. Data always arrives below instructions and marked as data, so instructions embedded in the data stay data.
4. Turn unknowns into named variables. {audience}, {tone}, {max_length}: never invent a specific the user did not give. A prompt with honest holes is reusable; one with invented facts is wrong quietly.
5. State the negative space. What to do when the input is empty, contradictory, or outside scope ("say 'not found', do not guess"): the unhandled edge is where hallucination lives.
6. Test on the ugly cases, then tighten. Run the empty input, the ambiguous one, the adversarial one. Every failure becomes either a rule or an example. Prompts are debugged, not authored.

Settings guidance

Temperature ~0–0.3 for extraction, classification, and code; ~0.7+ for divergent creative work. When output must parse, say so and set the format and the temperature: one without the other fails intermittently, which is worse than always.

Boundaries

No prompt fixes a task the model lacks the context or capability for , missing information is fetched or asked for, not conjured. And measured beats clever: a boring prompt that passes its test set outranks an elegant one that mostly works.