Prompt engineering questions, answered

A good prompt is specific, scoped, and output-shaped. Give the model a role, the goal, the audience, the format, and any constraints — in that order.

Minimum viable template: "You are [role]. Do [task] for [audience]. Return as [format]. Constraints: [list]."

1. Role — who the model is pretending to be.
2. Context — background the model needs.
3. Task — the one thing you want done.
4. Format — how the output should look.
5. Constraints — length, tone, what to avoid.
6. Examples — 1–3 samples of ideal output.

Length isn't the goal — clarity is. If cutting a sentence doesn't change the output, cut it. If adding one removes a common failure mode, keep it.

Replace "write a good product description" with "write a 60-word product description for a first-time buyer, in second person, ending with a one-line CTA".

Roles set vocabulary, tone, and priorities. A "senior editor" will trim; a "junior writer" will pad. Choose the role that matches the output you actually want.

One good example teaches the model more than a paragraph of instructions. Two examples show the pattern. Three cover edge cases. More than that usually adds noise.

Treat prompts like code: version them, test one change at a time, and keep a note of which change fixed which failure mode.

A useful pattern: "First list the steps you'll take. Then do step 1. Then do step 2…" You get better output and a debuggable trace.

Every prompt in this library is templated this way — the bracketed tokens are the only parts you edit before sending.

• Vague format — say "markdown table with columns X, Y, Z", not "nicely formatted".
• Contradictions — "concise but comprehensive" is a coin flip.
• Too many asks — one prompt, one deliverable.
• No negative constraints — tell it what not to do.

Full walk-through with tools, examples, and a copy-paste template: How to create an AI prompt from a photo.

Browse ready-made image prompts by tool on the Tools page — each app (Midjourney, Nano Banana, Flux, DALL·E) has its own prompt conventions.

Every prompt in the library that targets ChatGPT opens directly in a fresh chat with the template pre-filled — use the "Use here" button on any prompt page.

Logos need simplicity constraints more than style adjectives — "one color, no text, works at 32px" produces usable marks; "creative and unique" produces noise.

"Best" depends on the target model. A prompt written for Midjourney rarely works verbatim in DALL·E. Match the prompt writer to the destination tool.