Shaping the response
A model will happily answer in prose unless you tell it otherwise. Output formatting instructions specify the structure you want, such as a list, a table, a single word, or a fixed set of fields. Clear formatting makes downstream parsing far easier.
Techniques that work
- Name the format explicitly, for example respond with a bulleted list.
- Show a template with placeholders so the model fills it in.
- Constrain length by asking for one sentence or at most three bullets.
- Forbid extras by saying do not add explanation or preamble.
Common failure modes
- The model adds a friendly intro line you did not want.
- It wraps the answer in commentary that breaks a parser.
- It drifts from the requested fields over a long response.
Making it robust
Put format rules near the end of the prompt where they stay salient, and repeat the single most important constraint. When exact structure matters, pair the instruction with one example so the model has a concrete target to match.
Key idea
Output formatting instructions tell the model the exact shape of the answer, and naming the format plus showing a template keeps results parseable and consistent.