A prompt is a brief for a capable colleague who knows nothing about your situation. Most poor results come from underspecified prompts, not weak models.
The Building Blocks
- Context: who you are, what you're working on and why.
- Task: exactly what to produce.
- Audience: who will read the result.
- Constraints: length, tone, things to include or avoid.
- Format: bullets, a table, JSON, a specific template.
Before and After
Summarise this report.
versus
I'm briefing our finance director before a 10-minute meeting. Summarise the attached quarterly report in five bullet points, focusing on anything that affects next quarter's budget, and flag figures that look inconsistent. Plain language, no jargon.
Habits That Help
- Put long documents first and the question after, clearly separated (for example inside
<document>tags). - Say what to do, not only what to avoid.
- Explain why a constraint exists, so the model can make sensible judgement calls.
- Allow uncertainty: "If the document doesn't say, answer not stated."
- Iterate: treat the first answer as a draft and give specific feedback.
Common Mistakes
Vague goals, missing context the model can't know, several unrelated tasks in one prompt, and no description of what a good answer looks like.
Test Prompts Like Code
Try a prompt on a handful of realistic inputs, including awkward ones, before relying on it — and keep the versions that work.