Making AI Prompts Work Better, Step by Step (from Google)
Just read Google’s 68-page prompt engineering guide.
A few specifics you can use right away:
Start with a role + outcome:“You are a fashion stylist. Suggest 5 outfits for a summer wedding. Return JSON with fields: dress_code, items, why.”
Give one or two examples (few-shot):“Input: ‘outdoor, semi-formal’ → Output: {…}”Then your real input. This anchors format and tone.
Force step-by-step thinking (Chain-of-Thought without overdoing it):“Think in steps: 1) restate the goal, 2) list constraints, 3) propose 3 options, 4) pick the best and explain why (2 sentences). Return the final answer only.”
Ask for structure every time:“Return valid JSON with fields: summary, steps[], risks[], next_actions[]. No extra text.”
Improve reliability with self-check:“Before finalizing, verify: a) constraints met? b) sources cited? If any fail, revise once.”
Compare alternatives quickly (light self-consistency):“Generate two different approaches. Then pick the stronger one and explain the trade-off in one sentence.”
Control style and length:“Tone: concise, neutral. Max 120 words. Bullets if possible.”
For data tasks, pin the schema:“Output CSV with columns: id, title, category, confidence. No header notes.”
For ideation, add guardrails:“Give 10 ideas. Each must be distinct by audience or channel. No duplicates; skip clichés.”
For research summaries, require citations:“Summarize into 5 bullets. After each bullet, include a bracketed source like [Author, Year].”