TL;DR: Perplexity retrieves before it writes, so a research template's job is steering what gets found and forcing every claim to arrive with something you can check. Below: an evidence contract you paste into any prompt, plus 24 templates for literature scans, competitive work, due diligence, fact-checks, topic tracking and sourced briefs.
Why does a research prompt need a different shape from a chat prompt?
Because the same block of text is doing two jobs, and only one of them is writing. Perplexity's Agent API prompt guide opens by warning that prompts tuned for single-shot models often underperform "because the same text shapes tool selection, search query generation, and final response together" (docs.perplexity.ai, accessed August 27, 2026).
The mechanics of that inversion are the subject of a separate post, how to prompt Perplexity. This page assumes you have accepted the inversion and want the library.
What makes research prompts specifically hard is that the failure mode is quiet. A chat model that misunderstands you produces obviously wrong prose. A search product that retrieves the wrong pages produces well-formatted, confidently cited prose about slightly the wrong thing, and the citation panel makes it look verified. Perplexity's own hallucination guidance names the conditions: thin results, information that is not web-accessible, and repeated searches returning related but non-matching results.
Perplexity's research benchmark describes the task shape better than most prompt guides do. Its wide-research documentation defines the job as one where you "discover a broad set of items (wide) and investigate each far enough to back every claim with a source (deep)". Every template below is built for that second half. Discovery is the easy part.
What belongs in every research template?
Six slots and an evidence contract. The slots steer retrieval; the contract governs what comes back. You will not use all six every time, but the contract goes on every prompt without exception.
[ENTITY] The thing, named the way its own documents name it.
[CLAIM TYPE] What you want established, not what you want discussed.
[SOURCE CLASS] Which kind of source wins when two disagree.
[WINDOW] A date range, or a recency bound, in the text.
[SHAPE] Table, list of N, one paragraph per item. Say the number.
[NEGATIVE] What must be reported as absent if it is absent.
That sixth slot is the one nobody writes and the one that pays. A research answer that omits what it failed to find is not a research answer, it is a highlight reel. Here is the reusable block. Paste it at the end of any prompt on this page, or store it once as a Project instruction and stop pasting it.
EVIDENCE CONTRACT
For every factual claim, give me three fields in this order:
1. CLAIM - one sentence, no hedging.
2. SOURCE - publication name, domain, and publication date.
3. QUOTE - the sentence from that source that supports the claim,
reproduced exactly. Do not paraphrase it.
Rules:
- If you cannot produce all three fields, do not make the claim. Put it
under a heading "COULD NOT CONFIRM" with the search terms you tried.
- If results are related but do not match - a different year, a parent
company, a subsidiary, a similar product - state the mismatch before
you answer anything else.
- Do not write URLs into the prose. I will read them from the sources.
- Cap every list at [N] items, and say explicitly if you hit the cap.
- Separate what a source claims about itself from what a third party
measured or reported.
Two of those rules come straight from Perplexity's documentation rather than from folklore. Its prompt guide tells developers to instruct the model to say when a search returned nothing useful, and to "state the mismatch explicitly before answering" when results are near-misses. The rule about URLs comes from the same page, which warns that the model is "prone to mistyping or paraphrasing" URLs written into the answer text, and that the structured source list in the response is the authoritative one. In the app, the equivalent authoritative list is the sources panel.
Which constraints does Perplexity actually enforce?
Only the ones expressed as parameters, and in the consumer app there are almost none. This is the single most useful thing to understand before you trust a template that says "only use official filings".
Perplexity states the principle plainly: parameters "are applied by the search backend on every call, while prose-based filters are interpreted by the model and may not carry through every turn of the loop". So a research prompt has two kinds of constraint in it, and you should know which is which as you write.
| What you want | Enforced where | What to write |
|---|---|---|
| Only these domains | API only, search_domain_filter, up to 20 entries | In the app, name the domains in prose and check the sources panel afterwards |
| Published between two dates | API only, search_after_date_filter and search_before_date_filter | In the app, put the range in the query text and reject anything outside it by hand |
| Only scholarly sources | Sonar search_mode: "academic"; not listed on the Agent API filter table | In the app, name journals and say "peer-reviewed" in the query |
| Only SEC filings | Sonar search_mode: "sec"; not listed on the Agent API filter table | In the app, name the form types: 10-K, 8-K, DEF 14A |
| Nothing older than a month | API only, search_recency_filter | In the app, state the cutoff date explicitly rather than saying "recent" |
| A specific language | Sonar search_language_filter, up to 10 codes | Not listed on the Agent API filter table; profile setting in the app |
The practical consequence for the templates below: every one of them states its date window and source class inside the prompt text, because that is the only channel the app gives you. Treat those lines as a strong request rather than a guarantee, and spend the thirty seconds checking the sources panel.
Where should a research template live so you stop retyping it?
In a Project, in the Context settings. Perplexity's help centre describes that tab in two lines: "Add and manage instructions that are used for every query run in the project" and "Add and manage web links and domains to be prioritized". Instructions are capped at 8,000 characters (perplexity.ai help centre, article last updated July 30, 2026, accessed August 27, 2026).
That is exactly the right home for the evidence contract. Put the contract in the Project instructions once and every question you ask inside that Project inherits it, which means your typed prompt shrinks back down to the actual question. Note the published verb on the domain list, though: prioritized, not restricted. It is a weighting, not an allowlist, and the sources panel is still the check.
One more surface fact worth knowing before you design a template around a mode. Research mode "automatically selects the optimal models for your research needs, so you cannot manually choose a specific model", and Perplexity's newer Deep Research article adds that for broad queries "Research now asks clarifying questions before starting". A good template pre-answers those questions, which is why the ones below carry so many bracketed fields.
| Feature | Search | Research mode | Agent API |
|---|---|---|---|
| Runs many searches per question | Depends on plan and Pro Search | Documented as dozens of searches | Bounded by preset and max_steps |
| You pick the model | Yes, or leave it on Best | No, selected automatically | Yes |
| Asks clarifying questions first | Not documented | Yes, for broad queries | Not documented |
| Hard domain allowlist | Not published | Not published | search_domain_filter, up to 20 |
| Scholarly-source mode | Not published | Not published | search_mode academic, on the Sonar page |
| Exact publication date range | Prose only | Prose only | search_after / before_date_filter |
| Standing instructions per project | Project context, 8,000 chars | Project context, 8,000 chars | instructions parameter |
| Output lands in an editable file | Not documented | Yes, reports stream into a file | Sandbox files you download |
How do you run a literature scan without losing the citations?
By making the unit of output a study rather than a paragraph, and by asking for the negative result. A scan that returns twelve papers and no list of what it could not find has told you nothing about coverage.
1. LANDSCAPE SCAN
Map the published research on [QUESTION] since [YEAR].
Prefer peer-reviewed journal articles over preprints, and label any
preprint as such. One row per study: first author and year, sample or
dataset, method, primary finding, journal.
Cap at [12] rows. Then list, separately:
a) any study you found but could not confirm was peer-reviewed,
b) subtopics where you found no studies at all.
[EVIDENCE CONTRACT]
2. FIND THE DISAGREEMENT
Where does the published literature on [QUESTION] disagree with itself?
For each disagreement: state the two positions in one sentence each,
name the strongest paper on each side with its year and journal, and
quote one sentence from each showing the conflict.
Do not resolve the disagreement. Do not average the positions.
If you find only one position, say the literature appears one-sided
and name what you searched for.
[EVIDENCE CONTRACT]
3. TRACE A CLAIM TO ITS PRIMARY STUDY
The claim "[CLAIM, quoted as you saw it]" is widely repeated.
Trace it to the earliest primary source you can reach. Give me the
chain: the popular retelling, any intermediate summary or press
release, and the original study.
At each link, quote the sentence that carries the claim, so I can see
where the wording changed. If the chain breaks, say where.
[EVIDENCE CONTRACT]
4. REPLICATION AND RETRACTION CHECK
For [PAPER: authors, year, title], report:
- direct replication attempts, with outcome and year,
- published critiques or comments, with journal and year,
- any correction, expression of concern, or retraction notice.
If you find none of the three, say so explicitly and state which
sources you checked. Absence is the answer I need, not silence.
[EVIDENCE CONTRACT]
That last one matters more than it looks. If you are working at any scale, pair these with a workflow for summarising papers without losing citations, and treat the scan as the reading list rather than the reading. Broader prompt sets for the writing stage live in 30 AI prompts for literature review.
How do you research a competitor using only what it publishes?
By naming the source class as the company's own surfaces, and by making "not published" a legal answer. The most common defect in competitive research is a specification that came from a third-party blog and got attributed to the vendor.
5. PUBLISHED-FACTS-ONLY PROFILE
Build a profile of [COMPANY] from its own published surfaces only:
its website, docs, changelog, newsroom, and any regulatory filings.
Fields: what it sells, who it says it is for, published limits or
quotas, published availability or regions, and the date each page
was last updated if the page shows one.
Where a field is not published, write "not published". Do not
estimate, and do not substitute a third-party description.
[EVIDENCE CONTRACT]
6. CHANGE LOG ACROSS A SET
For each of [COMPANY A], [COMPANY B], [COMPANY C], list every
announcement made between [DATE] and [DATE].
Prefer the company's own blog, changelog, docs and filings over press
coverage. One line per announcement: date, one-sentence summary,
source domain.
Mark anything reported only by a third party as UNCONFIRMED BY VENDOR.
If a company published nothing in the window, say so.
[EVIDENCE CONTRACT]
7. CONTESTED MARKET NUMBER
Multiple figures circulate for [MARKET METRIC, e.g. 2026 market size
of X]. Find them all, do not reconcile them.
One row per figure: the number, who published it, publication date,
and the stated methodology or scope in their own words.
Then flag which figures trace back to the same underlying source, and
which are genuinely independent.
[EVIDENCE CONTRACT]
8. PACKAGING COMPARISON
Compare how [COMPANY A] and [COMPANY B] package [PRODUCT CATEGORY],
using only each company's own pricing and documentation pages.
Columns: tier name, the three headline limits, and what the page says
is excluded.
Any figure not on the page is "not published". Note the date you read
each page, because these change without notice.
[EVIDENCE CONTRACT]
If your competitive work is more strategic than factual, the analysis-stage prompts in 20 AI prompts for competitive analysis pick up where these leave off.
What does a due diligence prompt need that a normal prompt does not?
An explicit demand for negative findings, and a hard separation between the public record and commentary about it. Perplexity's own Deep Research documentation names "professional due diligence" as a benchmarked task, but a benchmark measures finding things, and diligence is largely about establishing that a thing is not there.
9. ENTITY FACTSHEET FROM THE RECORD
Build a factsheet on [ENTITY, with jurisdiction if known] from
official registries, regulatory filings and the entity's own
disclosures only.
Fields: legal name and any former names, registration number and
jurisdiction, incorporation date, registered address, stated business
activity, and the most recent filing you can find with its date.
Any field not found in an official source is "not found in the public
record", not blank and not inferred.
[EVIDENCE CONTRACT]
10. OWNERSHIP AND LEADERSHIP TRAIL
For [ENTITY], identify current directors, officers and any disclosed
beneficial owners, using registry filings and official disclosures.
For each person: name, role, the date the role started per the filing,
and the filing that shows it.
List separately any name that appears in press coverage but not in a
filing. Do not merge those two lists.
[EVIDENCE CONTRACT]
11. REGULATORY AND LITIGATION RECORD
Search for any enforcement action, regulatory finding, sanction, or
publicly docketed litigation involving [ENTITY] between [DATE] and
[DATE].
For each: the issuing body or court, the docket or case reference,
the date, and the stated status.
State plainly whether a matter is alleged, settled, or decided, and do
not characterise a settlement as a finding of wrongdoing.
If you find nothing, write "no matters found in sources searched" and
name the sources you searched.
[EVIDENCE CONTRACT]
12. RED-FLAG SWEEP WITH NEGATIVES
For [ENTITY], check each of the following and report the result even
when the result is nothing:
- registered address shared with many other entities,
- rapid changes of registered name or officers,
- auditor or accountant resignations,
- regulatory warnings or investor alerts naming the entity,
- discrepancies between claimed and filed financial figures.
Answer each line with FOUND plus evidence, or NOT FOUND plus what you
searched. An empty line is not an acceptable answer.
[EVIDENCE CONTRACT]
How do you fact-check a single claim?
By adjudicating one claim at a time and forcing the answer into a verdict plus evidence, rather than a discussion. The prompt that produces a paragraph of nuance has not checked anything.
13. SINGLE-CLAIM ADJUDICATION
Claim to check: "[CLAIM, quoted exactly as it was made]"
Return exactly this structure:
VERDICT - Supported / Contradicted / Unverifiable from sources
BEST SOURCE - the strongest source either way, with date
QUOTE - the sentence that decides it
AGAINST - the strongest source pointing the other way, or
"none found", with what you searched
WHAT WOULD SETTLE IT - one sentence naming the document that would
resolve this if you could reach it
No preamble. No summary paragraph.
[EVIDENCE CONTRACT]
14. STATISTIC PROVENANCE CHASE
The figure "[NUMBER, with the exact wording around it]" is being cited.
Find its origin. Report:
- the original publisher, publication and date,
- the population, sample or scope the figure actually describes,
- any change in wording between the original and the citation
you were given,
- whether the figure has been updated or superseded since.
If the trail ends at a source that itself cites no one, say so.
[EVIDENCE CONTRACT]
15. QUOTE VERIFICATION
Someone attributes this to [PERSON]: "[QUOTE]".
Find the earliest verifiable instance of these words. Report the
publication, date and context, and quote the surrounding sentence so
I can see whether the meaning survives being extracted.
If the wording differs from what I gave you, show both versions side
by side. If you cannot find it in any primary source, say the
attribution is unverified and stop.
[EVIDENCE CONTRACT]
16. DATE AND VERSION CHECK
For [PRODUCT / STANDARD / REGULATION], establish:
- the current version or edition, and the date it took effect,
- what the immediately previous version was,
- the source that states the current version, with its date.
Prefer the publisher's own version page or changelog over any summary.
If two of the publisher's own pages disagree, quote both and say they
disagree. Do not pick a winner.
[EVIDENCE CONTRACT]
Template 16 exists because vendors contradict themselves constantly, and a research tool asked to resolve a contradiction will usually just pick one. Perplexity's own help centre contradicts itself on which models are current across three live articles, so this is not a hypothetical. The usual reasons a model invents things apply here with a twist: retrieval gives the invention a citation, which makes it harder to spot rather than easier.
How do you track a topic over time without re-reading everything?
By asking for the delta rather than the state, and by giving the model the boundary of what you already know. A prompt that asks "what is happening with X" returns the same background every week.
17. WEEKLY DELTA BRIEF
Topic: [TOPIC]. I last reviewed this on [DATE].
Report only what is new since that date. Nothing published before it,
however relevant.
Format: one line per development, each with a date and source domain,
ordered oldest to newest. Cap at [10].
If nothing new was published, say "no developments found since [DATE]"
rather than summarising older material.
[EVIDENCE CONTRACT]
18. CHRONOLOGY BUILD
Build a chronology of [EVENT OR PROCESS] from [DATE] to [DATE].
One line per dated development: date, what happened in one sentence,
source domain.
Mark any development reported by only one outlet as SINGLE-SOURCE.
Mark any date that sources disagree on as DISPUTED and give both.
Leave gaps visible: if nothing is documented for a stretch, say so
rather than smoothing over it.
[EVIDENCE CONTRACT]
19. STATUS OF RECORD
What is the current, confirmed status of [ONGOING SITUATION] as of
today, [TODAY'S DATE]?
Structure:
CONFIRMED - the most recent established fact, with its date
REPORTED - claims made but not confirmed, each attributed
EXPECTED - anything with a published future date, e.g. a
deadline, hearing, or scheduled release
UNKNOWN - the questions nobody has published an answer to
Do not blend the first two sections.
[EVIDENCE CONTRACT]
20. WATCHLIST SWEEP
For each of [ENTITY 1], [ENTITY 2], [ENTITY 3], report any development
between [DATE] and [DATE] in these categories only:
[funding / leadership change / product launch / regulatory action].
One line per item, with date and source domain. If an entity had no
development in a category, write "none found" for that category.
I want the shape of the grid filled in, including the empty cells.
[EVIDENCE CONTRACT]
How do you turn a search session into a sourced brief?
By specifying the brief's structure before the research runs, so the model is filling a form rather than composing an essay. The structure is also what makes the brief reviewable by someone who was not in the session.
21. ONE-PAGE DECISION BRIEF
Decision to be made: [DECISION].
Produce a brief with exactly these sections, in this order:
1. The decision, restated in one sentence.
2. What the evidence supports, maximum [5] bullets. Each bullet
carries a claim, a source, and a quoted supporting sentence.
3. What the evidence does not settle, maximum [3] bullets.
4. What would change the answer: name the specific document,
dataset or disclosure.
5. Sources, listed once, with dates.
No recommendation section. I will make the call.
[EVIDENCE CONTRACT]
22. EVIDENCE TABLE
Question: [QUESTION].
Return a single table and nothing else. Columns:
Finding | Source | Date | Supporting quote | Confidence
Confidence is one of: Primary source, Secondary reporting, Single
source only, Contested.
Sort so that Primary source rows come first. Cap at [15] rows.
Below the table, list what you searched for and did not find.
[EVIDENCE CONTRACT]
23. TWO-SIDED BRIEF
Question: [CONTESTED QUESTION].
Build the strongest documented case for each side. For each side:
- the position in one sentence,
- the three strongest pieces of evidence, each with source and quote,
- the weakest point in that case, stated plainly.
Give both sides equal length. Do not conclude, do not label one side
mainstream, and do not describe a position as debunked unless you can
quote the source that debunked it.
[EVIDENCE CONTRACT]
24. HANDOFF BRIEF WITH GAP LIST
I am handing [TOPIC] to someone who knows nothing about it.
Produce:
- a 150-word orientation, no citations needed,
- the [8] facts they must know, each with source, date and quote,
- the [5] open questions, each with the specific source that would
answer it if it could be reached,
- a list of everything I asked about that returned nothing useful,
with the search terms tried.
That last list is the most important section. Do not omit it.
[EVIDENCE CONTRACT]
How do you audit the answer without redoing the research?
Read the sources panel before you read the prose, and spot-check the quotes. Three checks take about two minutes and catch the failure modes that matter.
First, count the domains. If a ten-row table cites four domains, you have a synthesis of four pages, not a survey. Second, open the two or three sources carrying the claims you most want to be true, and confirm the quoted sentence is actually on the page in those words. Third, read the "could not confirm" section first, because a suspiciously short one usually means the contract was ignored rather than satisfied.
Perplexity now marks some cited domains with a shield icon reading Government, Academic or Trusted, and its help centre is refreshingly clear about what that does not mean: "A label describes the website as a whole, not any single article or claim on it." The same page adds that "a label is never a substitute for reading the source yourself", and that "Perplexity aims for accuracy, but reviewing the original sources yourself remains the best way to build confidence in an answer."
The independent evidence supports that caution. In March 2025 the Tow Center for Digital Journalism tested eight generative search tools, pasting article excerpts and asking each tool to identify the headline, publisher, date and URL. They "ran sixteen hundred queries (twenty publishers times ten articles times eight chatbots) in total". Collectively the tools "provided incorrect answers to more than 60 percent of queries". The article names two per-tool figures in its text: Perplexity answered 37 percent of queries incorrectly, and Grok 3 answered 94 percent incorrectly. That 37 percent is the better of the two figures CJR names, and it still means more than a third of attributions were wrong.
That study is roughly eighteen months old and every tool in it has shipped new models since, so read the number as calibration rather than a current benchmark. One behavioural finding has aged better than the percentages: "Premium chatbots provided more confidently incorrect answers than their free counterparts", because they declined to answer less often. Paying more buys you fluency, not accuracy. That is the entire argument for the evidence contract.
What will break these templates first?
Not the prompt text. The surfaces underneath it, and one of them has a published date.
Every Sonar documentation page now carries a notice stating that "Sonar will be supported until September 27, 2026", with the Agent API as the replacement for both new projects and existing integrations. If you have automated any recurring research against Sonar, that is a calendar item, not a someday. The Agent API's own web search filter table does not list search_mode or search_language_filter, both of which are documented on the Sonar filters page, so an academic-mode or language-scoped pipeline needs checking rather than assuming.
In the consumer product, three names have moved. Spaces are now Projects, and the old help-centre URL still resolves while serving a page titled "What are Projects?". Focus modes are no longer documented for consumer accounts: on August 27, 2026, exactly one of Perplexity's 207 English help articles mentioned Focus mode, and only to say that "When you sign into an organization, it's not possible to use focus mode", pointing instead to a Choose sources picker. And the same feature is called Research mode in one help article and Deep Research in another.
Stop rewriting prompts. Start shipping.
Works with ChatGPT, Claude, Gemini, Grok, Midjourney, Ideogram, Veo3 & Kling. 5.0★ on the Chrome Web Store.
Create An AccountWhere a prompt tool actually helps here
We build a prompt generator. We are not a search engine, we do not index anything, and nothing on this page is a pitch to use us instead of Perplexity.
The narrow, real thing a prompt tool does for research work is structural. Every template above is the same skeleton with different variables: entity, source class, window, shape, negative, contract. That is precisely the shape of a saved prompt template with fields you fill, and it is why retyping these by hand degrades them. The bits that get dropped under time pressure are always the last two, which are the two that make the answer checkable. Our extension works on Perplexity, and it biases toward constraint-first structure rather than the persona-first shape a chat model wants. If your sources are documents you already have rather than the open web, NotebookLM's source-based templates are the better fit and we say so.
What no prompt tool can do is make a citation correct. It cannot reach a page behind a login, it cannot tell you that the one paper that would settle your question was never indexed, and it cannot stop a fluent answer from being wrong. It can only make sure that when you are wrong, you find out in two minutes rather than after you have shipped the memo. Keep the hallucination exit clause in every prompt, and read the "could not confirm" list first.