AI prompts that work: a practical checklist

Prompting is not a magic phrase. It is a clear brief. Most weak AI answers come from a vague request, missing context or an output format the model has to guess.

A simple structure

A reliable prompt usually contains five parts:

1. Role or perspective. "You are a careful copy editor" sets a stance, though it is not a guarantee.
2. Task. State the action: summarise, compare, rewrite, classify, extract.
3. Context. Paste the source, name the audience and explain the decision the answer supports.
4. Constraints. Give length, tone, reading level, what to avoid and which sources to use.
5. Output format. Ask for a table, bullet list, JSON, headings or a specific number of options.

You do not need all five every time. You do need the parts that remove ambiguity.

Show, do not just tell

Examples are often more effective than adjectives. Instead of "write in a friendly tone", include two sentences that sound right. Instead of "make it concise", say "under 120 words" and show a sample.

For classification or extraction, provide a few labelled examples. For formatting, provide a template with placeholders.

Put the source before the question

When you supply a document, put the document first and the instruction last. This keeps the task fresh and reduces the chance that the model answers from memory instead of the text.

For long sources, ask the model to quote the relevant lines before answering. That makes it easier to check whether the answer is grounded.

Ask for the reasoning you need, not a performance

"Think step by step" can help on some tasks, but it can also produce confident-sounding rationalisations. A better request is often: "List the assumptions, then give the answer" or "State what you are uncertain about."

For decisions, ask for alternatives and trade-offs rather than a single recommendation. For facts, ask for sources and a confidence level, then verify.

Iterate in small steps

A single giant prompt is hard to debug. Break the work into stages:

• first, outline;
• then, draft one section;
• then, critique against the brief;
• finally, revise.

This is slower per message but usually faster overall because you catch misunderstandings early.

Common mistakes

• No audience. The model defaults to a generic expert tone.
• No length. You get a essay when you wanted three bullets.
• No source. The model fills gaps with plausible guesses.
• Too many goals. The answer spreads thin across everything.
• No example of "good". The model guesses your taste.
• Blind trust. A polished answer is not a verified answer.

A reusable template

Task: [what to produce]. Audience: [who will read it]. Context: [source or situation]. Must include: [points]. Avoid: [tone, claims, formats]. Length: [limit]. Format: [structure]. If information is missing, ask one clarifying question before answering.

Use the template as a starting point, then delete what does not matter. The best prompt is usually the shortest one that still removes the ambiguity.

For more on how these systems generate text, read how AI models are trained and what an LLM is.