THE 10 BEST AI TECH RESEARCH PROMPTS

AI technology research prompts are where the gap between amateur and professional use is widest — and most dangerous. Used carelessly, a model will hand you a confident market size, a tidy competitive read, and a list of citations, some of which do not exist. Used well, the same model becomes a tireless research analyst that scopes the question, maps the landscape, synthesizes what you feed it, and surfaces the sources you then go and verify yourself. The difference is entirely in the prompt: whether it lets the model be the source of truth, or forces it to show its work and flag what it cannot stand behind. This library is ten professional research prompts, written out in full with no placeholders, plus the variable framework that makes them reusable and the sequence for chaining them into a finished research package.

This is a working resource. Every prompt below is complete and ready to paste; the only thing you add is your own specifics in the bracketed slots — and the discipline to check what comes back.

How these prompts are built

Every prompt here follows the same shape, and for research that shape carries an extra job. Each one opens with a role assignment that makes the model a specific kind of analyst, supplies the context of the decision the research must inform, imposes constraints that block the model's worst habits, and names the exact deliverables it must return. In research prompts, most of those constraints exist for one reason: to stop the model from fabricating. The single most important instruction you can give a research model is to separate what it knows from what it is inferring, and to never present an unverified source or an estimate as a confirmed fact.

The prompts are reusable because they run on a small set of variables. Replace these tokens with your own specifics before running any prompt — that one step lets a single prompt serve a market-entry study, an investment thesis, and a technical evaluation without rewriting it.

These tokens are intentional fill-ins, not unfinished sections — the controls on the instrument. The prompts are grouped into five phases that follow the real arc of research: frame the question, map the landscape, go to the primary sources, evaluate the technology, then pressure-test and synthesize. That is also the order to run them in. Working this way is a practical example of how large language models change knowledge work — they compress the mechanical parts of research while raising, not lowering, the premium on human verification.

Phase 1 — Frame the question

Most bad research is just a good answer to the wrong question. This phase forces you to decide what you actually need to know before you spend any effort finding out.

1. Research scoping and question designer

This prompt is the highest-leverage one in the set because it prevents wasted work downstream. It casts the model as a research lead and makes it convert a vague subject into a structured plan: the core question, the sub-questions that would resolve it, the evidence each requires, and — crucially — the assumptions most likely to be wrong, which the research should attack first. Run it before anything else on every project; it is beginner-friendly precisely because the structure does the strategic thinking for you.

You are a research lead at a technology intelligence firm, scoping a new research project.

CONTEXT
- Subject: [TOPIC / COMPANY / TECHNOLOGY].
- Decision this research must inform: [DECISION, e.g. "whether to build, buy, or invest"].
- Audience for the final output: [AUDIENCE, e.g. "an investment committee"].
- Time available: [TIMEBOX, or "not specified"].

TASK
Turn this vague subject into a structured research plan.

DELIVERABLES
1. The single core question the research must answer, in one sentence.
2. 5-8 sub-questions that, if answered, fully resolve the core question. Order them by importance.
3. For each sub-question: the type of evidence that would answer it (primary data, filings, expert input, technical benchmark) and where that evidence likely lives.
4. The two or three assumptions most likely to be wrong, which the research should attack first.
5. What "good enough to decide" looks like, so the research knows when to stop.

CONSTRAINTS
- Frame questions so they can be answered with evidence, not opinion.
- Prioritize the questions whose answers would most change the decision.
- Do not begin answering the questions yet - design the plan only.

Phase 2 — Map the landscape

With the question framed, you need the lay of the land: how big the field is, who is in it, and where the gaps are. These two prompts produce that map — and they are the first place fabrication tends to creep in, which is why both insist on labeling every number.

2. Market landscape and sizing mapper

This prompt defines a category, lays out its value chain, and sizes it from the bottom up rather than parroting a single top-down headline number. Its sharpest constraint is that every quantitative input must be marked as sourced or assumed, so the estimate is auditable and an assumption can never quietly pass as a fact. It works on any frontier model and pairs naturally with thinking about how generative AI is reshaping whole categories.

You are a market analyst mapping the landscape around [TECHNOLOGY / CATEGORY] for [AUDIENCE].

CONTEXT
- Category: [CATEGORY].
- Geography / segment of interest: [SEGMENT, or "global"].
- Why we are looking: [GOAL].

TASK
Map the market structure and size it from the bottom up.

DELIVERABLES
1. A plain-language definition of the category and its boundaries - what is in, what is out.
2. The value chain: who does what, from upstream inputs to end customer.
3. The main customer segments and the job each is hiring this technology to do.
4. A bottom-up market-size estimate (units x price, or customers x spend), with every input number labeled.
5. The three forces most likely to expand or contract this market over the next 3-5 years.

CONSTRAINTS
- Build the size estimate from stated inputs I can check, not from a single top-down number.
- Mark every quantitative input as [SOURCED] or [ASSUMED]; never present an assumption as a fact.
- If you lack a reliable basis for a number, give a range and say what would narrow it.

3. Competitive intelligence analyst

This prompt profiles the players, groups them by strategic type, and finds the white space no one owns well. It is built to resist the model's instinct to invent precise-sounding figures: it must separate what is publicly verifiable from what it is inferring, and write "unverified" rather than fabricate funding, customer counts, or revenue. The instruction to choose comparison dimensions tied to the actual decision keeps it from producing a generic feature checklist.

You are a competitive intelligence analyst profiling the players in [CATEGORY].

CONTEXT
- Our vantage point: [WHO WE ARE, e.g. "a potential entrant" / "an investor"].
- What we need to decide: [DECISION].
- Known competitors to include: [LIST, or "you identify them"].

TASK
Produce a structured competitive map.

DELIVERABLES
1. The 5-8 most relevant players, grouped by strategic type (incumbent, challenger, niche, adjacent threat).
2. For each: their wedge, who they serve, their apparent moat, and their most visible weakness.
3. A comparison table across the dimensions that actually matter for [DECISION] - you choose the dimensions and justify each in one line.
4. Where the white space is: the underserved segment or unmet need no current player owns well.
5. The competitor most likely to be underestimated, and why.

CONSTRAINTS
- Separate what is publicly verifiable from what you are inferring; label inferences as such.
- Do not invent funding figures, customer counts, or revenue. If you do not know, write "unverified" and note how it could be checked.
- Choose comparison dimensions tied to the decision, not a generic feature list.

Phase 3 — Go to the primary sources

This is the phase that separates real research from confident guessing, and it is where AI is most likely to betray you. The two prompts here are designed around a single assumption: the model's citations are leads to verify, never evidence to cite.

4. Primary-source finder and verifier

This prompt casts the model as a librarian who refuses to cite anything it cannot stand behind. Instead of producing a bibliography you might trust by accident, it returns each candidate source with an explicit confidence label — verified, likely, or unverified — and the exact search you should run to confirm it. It is the most important prompt in the library, because it converts the model from a fabrication risk into a search-and-verify assistant.

You are a research librarian who specializes in primary sources and refuses to cite anything you cannot stand behind.

CONTEXT
- Claim or topic I need sourced: [CLAIM / TOPIC].
- Acceptable source types: [e.g. filings, peer-reviewed papers, official statistics, company docs].

TASK
Identify the primary sources that would substantiate this, and be explicit about your confidence in each.

DELIVERABLES
For each source you propose, give:
- What the source is and who published it
- Exactly which part of my claim it supports or contradicts
- A confidence label: VERIFIED (highly confident this specific source exists as described), LIKELY (probably exists, must be checked), or UNVERIFIED (you are inferring it should exist)
- The search I should run to confirm it (database, query, or institution)

CONSTRAINTS
- Never present a source as real unless you are confident it exists; when unsure, label it and say so.
- Do not fabricate DOIs, URLs, titles, or author names. A described-but-unconfirmed source is acceptable; an invented citation is not.
- Prefer the original source over any secondary write-up of it.
- End with the single most authoritative source to check first.

5. Literature synthesizer

This prompt turns a stack of papers or notes into a decision-useful brief — but only from material you provide, which is the constraint that makes it trustworthy. It must say "not covered in provided sources" rather than fill a gap from memory, attribute every claim to its source, and flag disagreement instead of averaging conflicting findings into a false consensus. That makes it safe to use precisely because it cannot wander off your evidence.

You are a research analyst synthesizing technical literature for a non-specialist decision-maker.

CONTEXT
- Topic: [TOPIC].
- Material: [PASTE PAPERS / ABSTRACTS / NOTES - synthesize only what I provide].
- The decision this informs: [DECISION].

TASK
Synthesize the provided material into a decision-useful brief.

DELIVERABLES
1. The state of knowledge in plain language: what is well established, what is contested, what is unknown.
2. The strongest finding for and the strongest finding against the relevant position, each tied to the specific source it came from.
3. Methodological caveats that should raise or lower my confidence in these findings.
4. What the literature does NOT yet answer that matters for the decision.
5. A two-sentence bottom line a busy executive could act on.

CONSTRAINTS
- Synthesize only the material I provided. If something is not in it, say "not covered in provided sources" rather than filling the gap from memory.
- Attribute every claim to its source so I can trace it.
- Flag where sources disagree instead of averaging them into a false consensus.

Phase 4 — Evaluate the technology

Now the research turns from "what exists" to "does it actually work and will it last." These two prompts pressure the claims a technology makes about itself and separate a durable shift from a hype cycle.

6. Technology due-diligence evaluator

This prompt assesses whether a technology does what it claims by first stripping the marketing language away from the core technical assertion, then naming the assumptions that have to hold for the claim to be true. It reads maturity honestly — research-stage, emerging, or production-proven — and hands you the three questions that would most quickly expose a weak claim in a vendor conversation. It refuses to take benchmark numbers at face value, which is where most technical overclaiming hides.

You are a technical due-diligence lead evaluating whether [TECHNOLOGY / PRODUCT] does what it claims.

CONTEXT
- What it claims to do: [CLAIM].
- Our use case: [USE CASE].
- Our constraints: [BUDGET / STACK / TIMELINE / REGULATORY].

TASK
Assess the technology against its claims and our needs.

DELIVERABLES
1. The core technical claim restated precisely, separated from marketing language.
2. What has to be true for the claim to hold - the load-bearing assumptions.
3. The maturity read: research-stage, emerging, or production-proven, with the evidence that would confirm it.
4. The failure modes and limits most likely to bite our specific use case.
5. The three questions to ask the vendor or team that would most quickly expose a weak claim.

CONSTRAINTS
- Distinguish what the technology can demonstrably do from what is projected or promised.
- Do not accept benchmark numbers at face value; note what context a benchmark needs to be meaningful.
- If the claim depends on conditions we cannot meet, say so plainly.

7. Trend-versus-hype signal analyst

This prompt argues both sides of a trend at full strength before reaching a verdict, which is the only honest way to judge one. It ties the trend to first-order drivers — cost curves, adoption, unit economics, regulation — rather than sentiment, and it names the concrete leading indicators you could actually track over your horizon. The output is a calibrated verdict with a confidence level, not a vibe. This is the discipline that keeps research credible amid the noise of AI agents and every other fast-moving category.

You are an analyst whose job is to separate durable signal from hype in [DOMAIN].

CONTEXT
- The trend or claim under scrutiny: [TREND].
- Time horizon that matters to us: [HORIZON].
- What we would do differently if it is real: [DECISION].

TASK
Assess whether this is a durable shift or a cycle peak, and on what evidence.

DELIVERABLES
1. The strongest case that this is a real, durable trend - the underlying drivers, not the headlines.
2. The strongest case that it is overhyped or early - what the excitement is overlooking.
3. The leading indicators to watch: specific, observable signals that would confirm or kill the thesis over [HORIZON].
4. What would have to be true in 12 and 36 months for the bullish case to hold.
5. A calibrated verdict: durable / plausible-but-early / mostly hype, with your confidence and the main thing that could change it.

CONSTRAINTS
- Argue both sides at full strength before reaching a verdict.
- Tie the trend to first-order drivers (cost curves, adoption, regulation, unit economics), not sentiment.
- Name concrete indicators I could actually track, not vague "watch this space" advice.

Phase 5 — Pressure-test and synthesize

The last phase tries to break your own conclusion, then packages what survives into something a decision-maker can act on — and checks it one final time.

8. Red-team and steelman analyst

This prompt does the thing most research skips: it attacks your own conclusion before you commit to it. It first builds the strongest fair version of your thesis, then ranks the most serious ways it could be wrong and identifies the single cheapest test that would most threaten it. The honesty constraint matters — if the thesis survives scrutiny, it is told to say so rather than manufacture objections, which is what makes the exercise more than theater.

You are a red-team analyst hired to attack a conclusion before we commit to it.

CONTEXT
- Our current conclusion or thesis: [THESIS].
- The evidence we are relying on: [KEY EVIDENCE].
- What is at stake if we are wrong: [STAKES].

TASK
First steelman our thesis, then attack it as hard as the evidence allows.

DELIVERABLES
1. The steelman: the strongest, fairest version of our own thesis.
2. The three most serious ways it could be wrong, ranked by how damaging each would be.
3. For each: what evidence would confirm the failure, and whether that evidence is currently observable.
4. The disconfirming test - the single cheapest check that would most threaten the thesis.
5. A revised confidence level in the thesis after this scrutiny, with the reasoning.

CONSTRAINTS
- Do not strawman our position to make it easy to knock down.
- Attack the evidence and the logic, not motives.
- If, after honest scrutiny, the thesis holds up, say so - do not manufacture objections.

9. Executive briefing memo writer

This prompt compresses raw research into a decision memo, not a report. It leads with the recommendation, gives the three reasons that support it, states the strongest counterargument fairly, and surfaces the assumptions that would falsify the whole thing. The hard length limit and the "lead with the answer" rule force the clarity that a busy reader needs, and it carries the sourced-versus-estimate labeling all the way through to the final number.

You are a chief of staff turning raw research into a decision memo for [AUDIENCE].

CONTEXT
- Decision to be made: [DECISION].
- Research findings: [PASTE THE OUTPUTS FROM EARLIER PROMPTS / YOUR NOTES].
- How much time the reader has: [e.g. "two minutes"].

TASK
Write a tight decision memo, not a report.

DELIVERABLES (in this order)
1. Bottom line up front: the recommendation in two sentences.
2. The three reasons that most support it, each one line.
3. The single strongest argument against, stated fairly, and why it does or does not change the recommendation.
4. Key assumptions and what would falsify them.
5. The decision being requested and the next concrete step.

CONSTRAINTS
- Lead with the answer; never make the reader hunt for it.
- Every sentence must earn its place - cut anything that does not affect the decision.
- Quantify with [SOURCED] figures only; flag any number that is an estimate.
- Keep it under 400 words.

10. Claim and source fact-checker

This is the prompt that should run last, against your own draft, before anything circulates. It extracts every checkable claim and labels each as supported, unsupported, or suspect, then tells you the exact check that would confirm or refute it. Its governing instruction is to treat any precise figure, date, or quote as suspect until sourced — because precision is exactly where errors hide — and never to fill the gaps itself.

You are a fact-checker auditing a draft before publication or circulation.

CONTEXT
- Draft to audit: [PASTE THE TEXT].
- Standard: every factual claim must be traceable to a real, checkable source.

TASK
Extract and pressure-test every checkable claim in the draft.

DELIVERABLES
A table with one row per factual claim:
- The claim, quoted from the draft
- Type: fact, statistic, quote, or attribution
- Status: SUPPORTED (a real source is cited or it is plainly self-evident), UNSUPPORTED (no source and not self-evident), or SUSPECT (specific enough to be wrong and worth checking)
- For anything not SUPPORTED: the exact check that would confirm or refute it

Then list, separately, the claims that would most damage credibility if wrong - check these first.

CONSTRAINTS
- Treat any precise figure, date, or quote as SUSPECT until sourced; precision is where errors hide.
- Do not assert a claim is true unless it is genuinely self-evident; "sounds right" is not SUPPORTED.
- Do not invent the sources yourself - flag what needs checking, do not fill the gaps.

The research stack: chaining them into a workflow

The biggest upgrade is not any single prompt — it is running them in sequence and feeding each output into the next. A one-shot "research this company for me" prompt asks the model to scope, gather, evaluate, and conclude all at once, and it does every step shallowly while quietly inventing whatever it lacks. Stacking lets each step go deep and inherit the prior decisions, which mirrors how a real research team actually operates, from scoping memo to final brief.

Run them in this order, passing the relevant output from each step into the next:

1. Research scoping and question designer
2. Market landscape and sizing mapper
3. Competitive intelligence analyst
4. Primary-source finder and verifier
5. Literature synthesizer
6. Technology due-diligence evaluator
7. Trend-versus-hype signal analyst
8. Red-team and steelman analyst
9. Executive briefing memo writer
10. Claim and source fact-checker

By the time you reach the memo, the model is working from a scoped question, a sized market, a competitive map, verified sources, and a pressure-tested thesis. The fact-checker at the end then audits the whole package against a fixed standard. The result is research that hangs together and that you can defend, because verification was engineered into every handoff rather than hoped for at the end.

The one rule that makes AI research safe

Everything in this library rests on a single discipline: the model proposes, you verify. This is not caution for its own sake — it is a response to a measured failure mode. A large cross-model audit that checked tens of thousands of model-generated citations against scholarly databases found fabrication rates ranging from roughly 11% to 57% depending on the model, domain, and prompt, with the problem worse on newer and more specialized topics. A citation that looks perfect — plausible title, real-sounding authors, a well-formed DOI — can still point to a paper that was never written.

That is why these prompts are built to label, flag, and surface rather than assert, and why the source-finder and the fact-checker exist at all. Treat the model as a brilliant, fast, slightly unreliable research assistant: superb at structuring the work, generating options, and synthesizing what you give it, and never to be trusted as the final word on a fact. The verification step is not the tax you pay for using AI in research — it is the part that makes the research real. The same discipline underwrites credible technology analysis whether or not a model was involved.

The Bottom Line

Most people use AI to do their research, and they inherit its confidence along with its fabrications. The professionals use AI to structure the inquiry, map the field, synthesize the evidence they supply, and surface the sources they then verify themselves — which is a fundamentally safer and more powerful way to work. The prompts in this library are good on their own and far better in sequence, because the sequence builds a research package that holds together from question to conclusion. Copy them, fill in your variables, run them in order, and check what comes back. The model is the analyst. You are still the editor — and the editor is the one who decides what is true.