Good prompts are clear, constrainable, structured and evaluable. Below is a practical 4‑step method with a starter template.

1) Set the role and goal

“You are a senior product copywriter. Write a short headline…”

2) Give structure

Ask for lists, tables, steps, or JSON. The model performs better when output format is explicit.

3) Add constraints and examples

• Hard constraints: length, style, audience, brand terms, banned words
• Do/Don't: what to include/avoid
• Few‑shot: 1–2 high‑quality exemplars are enough

4) Iterate

Evaluate on a small sample (accuracy/completeness/actionability/harmful rate). Record failures and fixes to build a team template library.

Starter template

Role: {who}
Goal: {what}
Audience: {who for}
Constraints: {style, limits}
Output: {format}
Examples: {few-shot}

Common failures and fixes

• Overlong, unstructured prompts → split by factors; front‑load key points.
• Unstructured outputs → request lists/tables/JSON.
• Vague generalization → provide 1–2 positive/negative examples.
• Noisy context → keep only highly related snippets.

Further reading

Transformer Explained · Multimodal AI · Evals & Launch Gates