Good prompting is half technical, half editorial. Here are the patterns I keep coming back to.
Role + context + task
The most reliable structure I’ve found is: who you want the model to be, what context it needs, then what you want.
You are a senior product designer reviewing a mobile app for accessibility. Here is the screen description: [description]. List the top 3 accessibility issues and suggest fixes.
Chain of thought for complex reasoning
When accuracy matters, ask the model to think out loud before answering.
Before giving me the answer, walk through your reasoning step by step.
This adds tokens but dramatically reduces confident-sounding wrong answers.
Constraints as guardrails
Explicit constraints reduce output variance. Be specific about format, length, and what to exclude.
Summarise this article in exactly 3 bullet points. Each bullet should start with a verb. Do not include any statistics.
Few-shot examples
When tone or format is hard to describe, show instead of tell. Three examples are usually enough.
When these patterns fail
None of these are magic. Complex multi-step reasoning still goes off the rails. Factual accuracy on obscure topics remains unreliable. The model is a very good editor and synthesizer — treat it as that, not as a database.