Prompts for research and writing

AI assistants are fluent, which makes them good at drafting and summarizing, and risky when accuracy matters: they can state invented facts in the same confident tone as real ones. A good research or writing prompt therefore says who the text is for, what it may rely on, and what to do when the material runs out.

For research and summaries

• Give the sources. Paste the articles, notes or data, and ask the model to use only them.
• Ask for citations to your numbered sources, so every claim can be checked.
• Say what to do with gaps: "If the sources do not cover it, say so."
• Separate agreement from disagreement, and facts from interpretation.
• Check the result. Open the cited passages for anything you will repeat or publish.

<ROLE>
You are a careful research assistant writing for {{audience}}.
</ROLE>

<SOURCES>
Use only the sources below. Cite each claim with the source number in brackets, such as [2].
{{sources}}
</SOURCES>

<TASK>
Explain {{question}}. Separate what the sources agree on from where they disagree.
</TASK>

<FORMAT>
At most 300 words: a two-sentence answer, then short sections with headings.
</FORMAT>

<CONSTRAINTS>
If the sources do not answer part of the question, say so instead of guessing. Do not invent sources, numbers or quotes.
</CONSTRAINTS>

For writing

• Audience and purpose: who reads it and what they should do or know afterwards.
• Tone with an example: "plain and warm, like this paragraph: …", which works better than adjectives alone.
• Structure and length: headings, a word limit, what goes first.
• Your own material: key points, facts and quotes to use, and a rule against adding others.
• Words to avoid, such as superlatives you cannot prove or jargon your readers do not know.

Keep what works

A template like the one above can be reused for every question by filling in its variables; try the template filler. Before you rely on a prompt, run it through the prompt checker, and read what makes a good prompt for the general structure.