How to Write Prompts That Get Useful AI Answers

The five parts of a working prompt

Every reliable prompt contains the same skeleton, in roughly this order:

• Role: who the model should be for this task - "You are a senior backend engineer", "You are a concise technical editor". The role sets vocabulary, depth and what counts as a good answer.
• Task: one clear verb and one object - "rewrite this email to be firmer", "explain this error to a junior developer". One task per prompt; two tasks get two half-answers.
• Context: the facts the model cannot know - your situation, constraints, audience, what you already tried. This field is where most prompts are empty, and where the biggest quality gains live.
• Format: the shape of the output - bullets, a table, a 150-word paragraph, JSON with these fields. Without it, the model picks its default shape, which is rarely yours.
• Constraints: the limits - length, tone, what to exclude, what not to assume. "Under 200 words, no jargon, do not invent APIs" rules out the failure modes you know you will hit.

Why vague prompts produce generic answers

A model answers the prompt it receives, including the prompt it has to fill in for you. "Help me with my business email" has no audience, no purpose, no context - so the model answers the statistical center of every business email it has ever seen: a safe, middle-length, hedged paragraph that would be a bad draft for 90% of real situations.

The pattern generalizes: generic in, generic out, in proportion. Each field you leave empty gets filled with the average of your topic's training distribution. The practical rule is to treat every missing field as a hidden "average this" instruction - and either fill it or deliberately accept the average.

Vague:     "Write a description for my tool"
Specified: "Write a 2-sentence product description for a free
            in-browser PDF compressor. Audience: non-technical
            office users. Tone: plain, confident, no hype.
            Emphasize: no upload, no sign-up."

Examples beat adjectives

The single highest-leverage addition to any prompt is one concrete example of the output you want. "Make it punchy" asks the model to guess what punchy means to you; a two-line example of a sentence you like (and one you dislike, if the contrast matters) shows it. One good input-output pair calibrates style, length and register better than a paragraph of adjectives.

The same logic applies to structure: if you want a table, show the column headers; if you want JSON, show the schema. Models pattern-match hard, and your example is the strongest pattern in the prompt.

The iteration loop

When the output misses, the instinct is to rephrase the whole prompt. Better is a loop that edits one field at a time:

1. 1Identify exactly what missed: too long? wrong tone? missing a constraint you assumed the model would share? Name it in one sentence.
2. 2Add the missing constraint to the prompt - "under 100 words", "no introductory sentence", "second person" - and regenerate.
3. 3Keep what worked. Each iteration should accumulate constraints; a prompt that is rewritten from scratch every time never converges.
4. 4Stop when the output is good enough three times in a row - that is your reusable template for this kind of task.

The generator's job is to force the fields you would otherwise skip. Fill in role, task, context, format and constraints, export the prompt, and take it to whichever AI tool you use - the structure is what transfers, not the tool.

Frequently asked questions