How to prompt AI for research

By Dr. Rafiq Muhammad, MD, PhD · Updated June 2026

Most disappointing AI output is a prompting problem, not a model problem. A vague request gets a vague, generic answer; a structured one — with a role, your own material, a specific task, and clear constraints — gets something genuinely useful and, crucially, checkable. Here are the patterns that work, and the verification habit that has to come with them.

The role–context–task–constraints pattern

Four ingredients turn a wish into a brief:

• Role — who the model should act as (“a methodology reviewer in public health”).
• Context — the situation and any material to work from (your draft, your data, your notes).
• Task — the one specific thing you want done.
• Constraints — length, format, tone, and what to avoid.

Ground it in your own material

The single biggest upgrade: give the model your text, data, or notes to work from, rather than asking it to recall facts. Grounding it in real material keeps it relevant to your study and sharply cuts fabrication. The caveat is confidentiality — never paste sensitive, identifiable, or unpublished data into a tool that may retain it (your data management plan and ethics approval govern this).

Iterate, and ask it to show uncertainty

Treat it as a short conversation, not a one-shot. Start broad, then steer: “make it more concise,” “focus on the limitations,” “show your reasoning.” And explicitly ask the model to flag what it’s unsure about and to separate what it knows from what it’s inferring. That surfaces the soft spots you most need to check.

The habit that never goes away: verify

No prompt makes output trustworthy by itself. Even a perfectly engineered prompt produces claims, not facts — so every statistic, quotation, and citation still gets checked against the real source. Good prompting makes verification faster; it never makes it optional.