Prompting strategies that actually work

A practical checklist: structure, formatting, few-shot examples, chain-of-thought and grounding.

Structure every prompt in three parts

1. System — persona and rules (“You are a precise technical writer. Answer only from the context.”).
2. User — the concrete request.
3. Context — the source material to ground the answer.

Keeping these separate makes prompts predictable and easy to iterate on.

High-leverage techniques

• Be explicit about output format. “Reply with a JSON object” or “one sentence, no preamble”.
• Few-shot examples. Show 1–3 input→output pairs; the model mimics the pattern.
• Chain-of-thought. “Think step by step” improves multi-step reasoning (you can hide the steps from the user).
• Decomposition. Break a big task into smaller prompts and combine the results.
• Grounding. Provide sources and instruct “answer only from the context above; if unknown, say so.”

Anti-patterns to avoid

• Vague asks. “Tell me about X” invites rambling. Specify scope and format.
• Overstuffed context. Irrelevant chunks dilute attention — retrieve fewer, better chunks.
• Trusting unverifiable claims. For facts and math, prefer a tool over the model’s memory.

A quick checklist

• Is the persona/role set?
• Is the desired output format explicit?
• Is trusted context provided (and is it relevant)?
• Did I tell it what to do when it doesn’t know?

If an answer is bad, the fix is usually better context or a clearer format, not a longer question.