Research Assistant Prompts

Prompts for finding things out, and for not being fooled by what you find.

The rule that governs all of these

AI invents citations that look perfectly real: plausible authors, plausible journals, plausible years, plausible DOIs. This has ended careers and lost court cases. Open every source before you cite it. If it takes more than ten seconds to find, assume it does not exist.

Every prompt below is built to make verification easy rather than to sound authoritative.

Starting a topic

The lay of the land

I am researching [TOPIC] for [PURPOSE].

Give me:
1. The three to five positions people actually hold on this
2. What each side's strongest evidence is
3. Where the genuine uncertainty is, versus where it is settled
4. The vocabulary I need to search effectively
5. Who the serious voices are

Label clearly what is well established, what is contested, and what you
are uncertain about.

The question refiner

I want to find out [VAGUE QUESTION].
That is too broad to research well. Break it into five specific,
answerable sub-questions, ordered so earlier answers inform later ones.
Tell me which one I should start with.

Finding sources

Citation-first search (use with a tool that has live web access)

Find sources on [QUESTION]. For each:
- Full citation
- A direct link
- One sentence on what it actually claims
- Its date
- Why it is credible, or why it might not be

Only include sources you can link to. Do not describe a source you cannot
point me at. If you find fewer than five, return fewer than five.

The verification pass

Here is a list of sources: [PASTE]

For each, tell me:
- Can you confirm this exists? How?
- If you cannot confirm it, say so plainly.
Do not guess at a DOI, volume or page number.

Reading critically

The document interrogation

I have given you [DOCUMENT]. Do not summarise yet.

1. What is the central claim?
2. What evidence is it based on?
3. What does it assume without arguing for?
4. Who funded or produced it, and how might that shape it?
5. What would a serious critic say?
6. What does it not address that matters?

The quote check

For each claim in your summary, quote the exact sentence from the source
that supports it. If you cannot find a supporting sentence, mark the claim
NOT SUPPORTED rather than rewording it.

This one catches more errors than anything else on this page. Summarising drifts; quoting cannot.

Comparing sources

Here are [N] sources on the same question: [PASTE OR UPLOAD]

Build a table: claim | which sources support it | which contradict it | which are silent.
Then: where do they genuinely disagree, and what would settle it?
Do not average them into a consensus that none of them holds.

Synthesis

The structured brief

Write a research brief on [TOPIC] for [AUDIENCE].

Structure:
- What we know with confidence, and how we know it
- What is contested, and what the disagreement turns on
- What we do not know
- What this means for [DECISION]
- What I should verify before acting

Mark every claim: [ESTABLISHED] / [CONTESTED] / [UNVERIFIED].

The honest gap list

Based on everything we have covered, list what I still do not know.
Separate: things I could find out with more research, things nobody knows,
and things I would need primary data to answer.

Domain-specific

Academic

Explain the methodology of this paper in plain English.
Then: what could go wrong with it? What would a reviewer push back on?
Is the sample adequate for the claim? Does the conclusion match the data,
or does it reach further?

Market and competitor

Research [COMPANY/MARKET] for [PURPOSE].
Distinguish clearly between: their own marketing claims, independent
reporting, and your inference. Label each.
Give me the date of anything time-sensitive — a two-year-old funding
figure is not current.

Technical evaluation

Compare [TOOL A] and [TOOL B] for [USE CASE].
Cover: what each is genuinely good at, the failure mode of each,
switching cost, and who each is actually built for.
Flag anything where your information might be out of date.

The verification protocol

Before anything researched with AI leaves your hands:

• Every citation opened and confirmed to exist
• Every citation actually says what it was claimed to say
• Every number traced to a source
• Every date checked for currency
• Contested claims labelled as contested
• Anything I could not verify is marked, not quietly dropped
• A human who knows the field has looked at it, if it matters

When to stop using AI for this

• When the answer must be right and you cannot verify it yourself.
• When the field is specialised enough that you cannot tell a plausible wrong answer from a right one.
• When the material is behind a paywall the AI cannot reach — it will summarise the abstract and present it as the paper.
• When you need primary data. AI cannot run your survey.