ChatGPT, Gemini, Claude and Perplexity each have a research mode that searches the web on its own and returns a long, cited report. These briefs are written so the run needs no input from you once it starts, and the report still has to be checked before you rely on it. These prompts cover both ends of the run.
There are 47 prompts in eight groups. Three groups write the brief: interviewing you first, a one-shot five-part brief, a context dump and a pre-flight check for guesswork; briefs for science, history and policy questions, peer comparison, customer expectations, identifying something from fragments, products against hard specs and trend scans; and source rules, ending with an appendix that records what the run could not read. Two more groups set the report format and take the report to a decision.
The checking group works on what the tool returned: its claims, its cited source list, its numbers and its source dates, plus a known-answer test, a comparison with a plain answer and a comparison of two tools' reports. The briefs and checks are written to work across all four tools. A separate group of five prompts uses features that Google, Anthropic and Perplexity describe on their own help pages, and the note under each says which.
Reports can still be wrong, so these prompts push you towards opening the original sources.
Write the Brief
A deep research run works on its own once it starts, so the brief has to answer the questions before the tool can ask them. These prompts interview you first, write a complete five-part brief, add a hypothesis that invites counter-evidence, set a stopping rule that allows 'not found', split big questions into separate runs, pack your background in as context rather than evidence, and check a draft brief for anything that would force the tool to guess.
Best for: Getting a complete brief before a long run
I am about to start a deep research run. Be the clarifying step the research tool may skip, then write the brief.
My starting request: {{STARTING_REQUEST}}
What I will use the result for: {{INTENDED_USE}}
Step 1. Ask me at most 7 numbered questions, most important first, each answerable in a sentence (offer 2 or 3 options where useful). Cover only what my request leaves open from this list: purpose, audience, scope, time window, region, sources I trust or distrust, and the report format I want. Then stop and wait for my answers.
Step 2. After I reply, write one brief I can paste into the research tool, using these labelled parts: Purpose, People (who reads it and what they know), Process (what to explore, what is allowed and what is not), Platform (sources to use or avoid), Performance (the output and its structure).
Once the run starts it will work on its own, so end the brief with a line "If unclear" giving the default the tool should follow for each open point. Mark every detail I did not give you as "TO CONFIRM" instead of filling it in. Keep the brief under 300 words.
Fill in: STARTING_REQUEST INTENDED_USE
Use in an ordinary chat first, then paste the brief into the research mode. The five labels follow the Purpose / People / Process / Platform / Performance framework in Trust Insights' deep research prompting guide.
Best for: Avoiding confirmation bias in the brief
Draft a deep research brief that tests my hypothesis instead of confirming it.
My hypothesis: {{HYPOTHESIS}}
Context: {{SITUATION}}
Why I might be wrong, in my own words: {{MY_DOUBTS}}
The brief should instruct the research tool to:
1. State the hypothesis neutrally, as a question.
2. Search separately for evidence that supports it and evidence that contradicts it, and give both equal space.
3. Report what would count as strong versus weak evidence either way, and which kind each source provides.
4. Say plainly if the evidence is mixed, thin or absent. An honest "unclear" is a valid conclusion.
5. End with the single most important unresolved question.
Write the brief in direct instructions the tool can follow, with no persuasive framing, under 250 words. Do not research the hypothesis yourself. After the brief, list the two ways my wording of the hypothesis could bias the result, and a suggested neutral rewording.
Fill in: HYPOTHESIS SITUATION MY_DOUBTS
Use when you suspect you already believe the answer.
Best for: Forcing honest gaps instead of filler
Help me write a stopping rule for a research run, so the report says when the evidence is insufficient instead of padding.
Question: {{QUESTION}}
The decision this feeds: {{DECISION}}
How costly it would be to be wrong (low, medium or high, and why): {{COST_OF_ERROR}}
Draft an instruction block for the brief that sets:
- The minimum evidence I need before the report states a finding as established (for example, how many independent sources, and what type).
- What the report must do when it finds only one source, only a vendor or promotional source, or only old sources: label the finding "weak" and say why.
- What the report must say when it cannot find an answer at all: state "not found", list where it looked, and name who or what might hold the answer.
- A ban on filling gaps with plausible-sounding general statements.
The run will not be able to check with me, so every rule must say what the tool does on its own. Scale the strictness to the cost of error I gave. Explain in two lines why you chose that level. Output: the instruction block, ready to paste.
Fill in: QUESTION DECISION COST_OF_ERROR
Adds honest 'not found' behaviour to any brief.
Best for: Breaking a broad topic into runs
A single deep research run on a very broad question tends to give a thin report on every part. Help me split mine into separate runs.
Big question: {{BIG_QUESTION}}
Decision it feeds: {{DECISION}}
Number of separate runs I am willing to do: {{NUMBER_OF_RUNS}}
Break the question into that many self-contained research questions. For each run give:
- A one-sentence question that could be answered without the other runs.
- What the run should deliver, in one line.
- What it depends on from another run, if anything, so I can order them.
- A two-line brief I can paste straight into the research tool.
Then add a short "how the pieces fit together" note: what I will do with the outputs once all runs are back. Do not answer any of the questions yourself.
Fill in: BIG_QUESTION DECISION NUMBER_OF_RUNS
Check your plan's current allowance on the vendor's help page before choosing how many runs to do.
Best for: Complete briefs for tools that don't ask
Run a research report. I cannot answer follow-up questions, so everything you need is below.
PURPOSE (why I need this): {{PURPOSE}}
PEOPLE (who will read it and what they already know): {{READER}}
PROCESS (the question, including any place or exclusions, then the time window; say what is allowed and not allowed rather than step-by-step instructions):
Question: {{QUESTION}}
Time window: {{TIME_WINDOW}}
PLATFORM (sources to prioritise, sources to avoid): {{SOURCE_RULES}}
PERFORMANCE (output format): {{OUTPUT_FORMAT}}
Requirements:
- Follow the output format exactly, with headers for each section, and put inline citations next to the claims they support.
- End with a list of all sources: title, publisher, date, link.
- For scientific or academic points, link directly to the original paper or official publication where possible.
- If any detail above is missing or unclear, do not invent it. State the gap and your assumption at the top.
- Mark claims that rest on one source only.
Fill in: PURPOSE READER QUESTION TIME_WINDOW SOURCE_RULES OUTPUT_FORMAT
OpenAI's developer guide says deep research through its API has no clarification step, expects a fully formed prompt, and advises stating the output format and priority sources. The five labels follow Trust Insights' Purpose / People / Process / Platform / Performance checklist.
Best for: Giving background without biasing the run
Turn my background material into a context block for a deep research brief. The tool must treat it as context, never as a source.
Research question: {{QUESTION}}
My notes and background (pasted as they are): {{BACKGROUND_NOTES}}
An earlier AI answer on this topic, if I have one: {{EARLIER_ANSWER}}
Write a block with these headings:
1. Context: the facts about my situation that shape the question, restated neutrally and briefly. Nothing from the notes that is not relevant.
2. Claims to verify: statements in my notes or the earlier answer that are about the outside world and sound factual, each listed so the research tool is asked to check it rather than assume it.
3. Do not cite: an instruction that the tool must not cite my notes or the earlier answer as sources, and must find its own.
4. Conflicts: an instruction that where its findings contradict my context, it says so with sources.
Keep it under 300 words. If my notes are too vague to be useful, say what to add.
Fill in: QUESTION BACKGROUND_NOTES EARLIER_ANSWER
Keeps your assumptions from being echoed back to you as findings.
Best for: Checking the brief before a long run you won't need to supervise
I am about to paste the brief below into a deep research mode. Once the run starts it will search and read on its own, and I will not be able to answer its questions or redirect it. Check the brief for anything that would force the tool to guess.
The brief: {{DRAFT_BRIEF}}
What I will do with the report: {{INTENDED_USE}}
List each problem with the exact wording affected:
1. Terms the tool could read two ways (for example "recent", "small firms", "the market"), with a suggested definition for each.
2. Missing scope: time window, region, what is out of scope.
3. Missing source rules: which sources count, which to avoid, how to cite.
4. A missing output format, or one that does not suit my intended use.
5. Any instruction to check with me, wait for me or let me choose. Rewrite each as a default the tool follows by itself and records at the top of the report.
6. Instructions that conflict with each other.
Then give the corrected brief in full, changing only what your list covers. Mark anything you had to assume "ASSUMED" so I can change it before I start the run.
Fill in: DRAFT_BRIEF INTENDED_USE
Run in an ordinary chat before you start the research. OpenAI's developer guide says its API deep research has no clarification step and expects a fully formed prompt; a brief that needs no questions suits every tool.
Briefs for Specific Kinds of Question
A good brief differs by subject, and so does where the run should search first. These cover science and health (an evidence map by study quality), history (primary sources and changing interpretations), policy and law (facts versus values), peer-set comparison, customer expectations, identifying something from fragments, spec-constrained product research and trend scans with independent-source counts.
Best for: Briefing science and health questions
Deep research brief for a scientific, medical or technical question. Search widely, read the studies themselves where you can, and organise what you find as an evidence map rather than one narrative.
Question: {{QUESTION}}
Why I am asking: {{PURPOSE}}
Time window for the evidence: {{TIME_WINDOW}}
Kinds of source to prioritise: {{SOURCE_TYPES}}
Organise the report into four parts:
1. Well established: findings supported by several independent studies or reviews.
2. Contested: points where credible sources disagree, with each position and the evidence it relies on. Do not smooth a disagreement into one story.
3. Thin or preliminary: findings resting on one study, a preprint, early data or a press release.
4. Not found: parts of the question the sources do not answer.
For each study you rely on, give its design (for example randomised trial, cohort, survey, lab study, expert opinion, vendor report), the sample size and type, whether it has been replicated, its main limitation and its date. Prefer systematic reviews, meta-analyses and official guidance to news coverage, and link the original paper rather than a write-up of it. Where you could read only an abstract, say so next to the study.
If the question concerns health, treat the report as background reading, not advice, and list what a qualified professional should confirm for my situation.
Fill in: QUESTION PURPOSE TIME_WINDOW SOURCE_TYPES
Adapted from the 'verifiable topics' template in Agents Today #12 (consensus versus contested versus speculation, study design, replication). OpenAI's developer guide advises linking directly to the original paper for academic queries.
Best for: Historiography-aware research
Deep research brief for a historical question. Search digitised archives, official records and scholarly work, not only encyclopaedia-style summaries, and give weight to evidence over narrative.
Question: {{QUESTION}}
Period: {{PERIOD}}
Place: {{PLACE}}
Aspect I care about most (political, economic, social, cultural): {{ASPECT}}
The report should:
1. Separate what primary sources record (documents, records, contemporary accounts) from what later historians conclude. Name which is which for each key claim.
2. Say who wrote each primary source, when, and what bias or limit that creates.
3. Describe how historians' interpretations have changed over time and the main disagreements, naming the positions.
4. Flag claims that appear widely repeated but trace back to one source, or that you could not trace to a source at all.
5. End with open questions the surviving evidence cannot settle.
Cite sources with author, title, date and link where available. Do not smooth a disputed event into one story, and do not attribute motives without a source.
Fill in: QUESTION PERIOD PLACE ASPECT
For history questions, where retellings are common and primary evidence is thin.
Best for: Briefing contested policy questions
Deep research brief for a policy, legal or ethical question. Search the positions of each side separately, so the first one you find does not set the terms for the rest.
Question: {{QUESTION}}
Jurisdiction: {{JURISDICTION}}
Perspective I currently lean towards: {{MY_LEAN}}
The report must:
1. Present the main positions, each described fairly enough that someone holding it would agree, with who holds it and the strongest evidence and strongest objection for each.
2. Mark clearly which statements are empirical (can be checked against evidence) and which are normative (depend on values). Do not treat a value judgement as a fact.
3. For empirical claims, cite the data or study and say how strong it is.
4. Identify competing frameworks or principles that lead to different conclusions.
5. Say where sources have an interest in the outcome.
6. Point out where evidence is thin, contested or missing.
7. Where the law or rules are involved, cite the legislation, regulator or court source directly and give its date.
Do not conclude which side is right. If my stated lean means a perspective is easy for me to underrate, name it. This is background research, not legal advice.
Fill in: QUESTION JURISDICTION MY_LEAN
Pairs an empirical/normative split with a dated legal-source rule. Verify rules with the official source before relying on them.
Best for: Finding something you only half remember
Deep research brief: I am trying to identify something I only remember in fragments. Search on each fragment separately before combining them, because one wrong fragment can hide the right answer. Then show me the evidence.
What I remember (as I recall it, including anything I am unsure about): {{FRAGMENTS}}
What kind of thing I think it is (product, game, book, place, event, other): {{CATEGORY}}
Roughly when I encountered it: {{WHEN_ENCOUNTERED}}
Where I encountered it: {{WHERE_ENCOUNTERED}}
Return:
1. Up to 5 candidates, ranked by how well they match my fragments.
2. For each, a table of my fragments against the candidate: matches, does not match, and unknown, with the source for each match.
3. The detail that would best tell the candidates apart, so I can check it.
4. Fragments that no candidate explains, which may mean my memory is wrong.
If nothing fits well, say so and suggest what to ask or where to look next. Do not claim a match based on one fragment, and do not invent a candidate to fill the list.
Fill in: FRAGMENTS CATEGORY WHEN_ENCOUNTERED WHERE_ENCOUNTERED
Zapier's guide shows the tool tracing down details like a half-remembered product or game; this prompt makes it show its matching.
Best for: Finding products that really fit the specs
Deep research brief: find products that meet my hard requirements and show how each one meets them. For every product, check the manufacturer's specification page for each hard requirement, then look for at least one independent review or test.
Product type: {{PRODUCT_TYPE}}
Hard requirements: {{MUST_HAVES}}
Nice-to-haves, in priority order: {{NICE_TO_HAVES}}
Budget: {{BUDGET}}
Country where I will buy: {{COUNTRY}}
Return:
1. A main table of at most 8 products for which you verified every hard requirement. Columns: product, each hard requirement with its evidence (source link), price with the date it was seen, and each nice-to-have.
2. A near-miss list: products that fail a hard requirement (name it) and products where you could not verify one, marked "unverified" with the requirement named. No unverified product goes in the main table.
3. A note on which claims come from the manufacturer or seller and which from independent reviews or tests.
Do not treat star ratings as evidence of a specific feature. If prices differ by seller, give the range and its sources.
Fill in: PRODUCT_TYPE MUST_HAVES NICE_TO_HAVES BUDGET COUNTRY
Good for consumer or equipment research where specs and prices change.
Best for: Spotting trends without hype
Deep research brief: scan for emerging trends in the area below and show how well each trend is evidenced. Search each source type I list separately, so one type does not crowd out the rest.
Area: {{AREA}}
Period to cover: {{PERIOD}}
Source types to scan (for example trade press, research reports, company announcements, filings, conference programmes): {{SOURCE_TYPES}}
Return a table of trends with columns: trend, what the evidence shows, number of independent sources that mention it, earliest and latest source dates, who is promoting it (vendors, analysts, independent researchers) and your assessment of whether it is early signal, growing, or hype.
Rules:
- Count a source once per publisher. Do not count articles that repeat the same press release as independent.
- Include trends that only a few sources mention, flagged as "weak signal".
- Show one source for each judgement.
- Add a section "What the scan may be missing" covering source types or regions it did not cover.
Fill in: AREA PERIOD SOURCE_TYPES
Source counts make it easier to tell trends from repeated marketing.
Best for: Benchmarking against similar organisations
Deep research brief: compare how similar organisations handle one activity. For each candidate peer, search its own published pages before relying on coverage of it.
My organisation, in a sentence: {{MY_ORGANISATION}}
Activity to compare: {{ACTIVITY}}
How to choose peers (size, sector, type): {{PEER_DEFINITION}}
Region: {{REGION}}
Return:
1. The peer set you chose, with the reason each organisation counts as a peer. If you could not find enough genuine peers, say so rather than stretching the definition.
2. A table: peer, what they do, how they describe it, source link and date. Use "not publicly found" where it is not published.
3. Patterns across peers, strengths, gaps and unusual approaches, each tied to the peers it is based on.
4. Where my organisation would stand against the peers, only as far as my description above allows. Mark what you would need to know about my organisation to say more.
Use the peers' own published material and independent coverage, and label which is which. Do not assume a practice is common because one peer does it.
Fill in: MY_ORGANISATION ACTIVITY PEER_DEFINITION REGION
Adapted from the peer-comparison style of university deep research examples (SDSU's AI brand library).
Best for: Separating survey evidence from anecdote
Deep research brief: find out what a group of people expect from a type of product or service. Search for surveys and studies first, then reviews and forums, then vendor material, and keep them apart.
Audience: {{AUDIENCE}}
Product or service type: {{OFFERING}}
Region: {{REGION}}
Time window: {{TIME_WINDOW}}
Source rules:
- Separate three kinds of evidence and label each claim with its kind: SURVEY OR STUDY (give the publisher, the date, the sample size and the method if stated), REVIEWS AND FORUMS (anecdotal; say how many you saw and where), and VENDOR OR CONSULTANCY MATERIAL (flag that the publisher sells to this market).
- Do not write that "customers want" something unless a survey or study supports it. If only reviews or forum posts do, say "reported in reviews".
- If a survey does not state its sample or method, say so.
Return expectations grouped by theme, each with its evidence kind, source and date, then a list of themes where sources disagree, and a list of what the research could not establish.
Fill in: AUDIENCE OFFERING REGION TIME_WINDOW
For market research where opinion pieces and real surveys otherwise blur together.
Set the Source Rules
Tell the run where to look and how to cite. These blocks set a source hierarchy with date rules, primary sources only, a closed list of sites and documents, publisher and conflict-of-interest flags, a separate label for your own files, and an appendix recording how the run searched and what it could not read.
Best for: Setting explicit source standards
Write a source-rules block I can paste into a deep research brief.
Field or topic: {{FIELD}}
Source types I trust most, in order: {{TRUSTED_SOURCE_TYPES}}
Source types or sites to avoid: {{SOURCES_TO_AVOID}}
Oldest publication date I will accept for factual claims: {{DATE_CUTOFF}}
How to treat figures that change over time: {{CHANGING_FIGURES_RULE}}
The block must tell the tool to:
1. Prefer the trusted source types in the order I gave, and use lower-ranked sources only to fill gaps, labelling them as such.
2. Not use the sources I want to avoid. If the only source for an important point is one of those, mention the point and say it rests on an avoided source rather than hiding it.
3. Not use sources older than my cutoff for factual claims. If an older source is the only one for a point, put the point in a separate list headed "Older sources only" with the source date, not in the findings.
4. For every factual claim, give the source name and date next to it. Where a figure changes over time, give it with its date and say if you could not find a current one. Mark undated sources "undated" and treat them as low-weight.
5. Mark any claim that rests on a single source.
Keep the block under 200 words and write it as instructions to the tool.
Fill in: FIELD TRUSTED_SOURCE_TYPES SOURCES_TO_AVOID DATE_CUTOFF CHANGING_FIGURES_RULE
Paste into the brief under a heading such as 'Source rules'.
Best for: Getting links to originals, not summaries
Deep research brief with a strict rule on sources. Follow links from secondary coverage back to the original before you cite anything.
Topic: {{TOPIC}}
Kind of claim I need supported (for example regulation, official statistics, company results, published studies): {{CLAIM_TYPE}}
Region: {{REGION}}
Period: {{PERIOD}}
For each claim you make about the topic, link to the original, primary source: the law or regulator's page, the statistical agency's release, the company's own filing or announcement, or the original paper or official journal publication. Do not cite a news article, blog or summary page when the primary source exists. If you can only find a secondary source, say "secondary source only", name it, and say what primary source I should look for.
Output:
- A table with the columns: claim, primary source (name and link), date of source, and note.
- Below the table, a list headed "Could not reach a primary source" with the claims affected.
Do not state a figure or rule from memory. If you cannot find a source, say so. Keep your commentary to what the sources say.
Fill in: TOPIC CLAIM_TYPE REGION PERIOD
OpenAI's developer guide advises linking directly to the original paper or official publication for academic or scientific queries; this prompt applies the same rule to other claim types.
Best for: Closed-universe research
Do this research using only the sources I name below.
Question: {{QUESTION}}
Websites or domains you may use: {{ALLOWED_SITES}}
Documents I have attached or connected: {{MY_DOCUMENTS}}
Rules:
- Use only the sites and documents listed. Do not use other websites even if they look more relevant.
- If the allowed sources do not answer part of the question, write "Not covered by the allowed sources" for that part. Do not fill the gap from general knowledge.
- Cite each point with the site or document name, and for documents the section or page where you found it.
- At the end, list which of the allowed sources you actually used and which you could not access or found irrelevant.
Format: a short answer first, then findings by sub-question. If your tool has a setting to turn off general web search or to restrict sites, I will have set it before running this.
Fill in: QUESTION ALLOWED_SITES MY_DOCUMENTS
Google's help page says to deselect Google Search to limit research to other selected sources, and to use Add Files to upload files. Use whatever source controls your tool offers, then keep this wording as a second guard.
Best for: Seeing who is behind each claim
Deep research brief: research the question below and be transparent about who is speaking. Look for independent sources as well as the interested parties' own material.
Question: {{QUESTION}}
Parties with a commercial stake in the answer that I know of: {{INTERESTED_PARTIES}}
For every major finding, tell me:
- Who published the source (independent researcher, regulator, trade body, vendor, advocacy group, media outlet, individual).
- Whether the publisher benefits commercially or politically from the claim. Flag the sources from the interested parties I listed, and any others you spot.
- Whether at least two independent sources support the finding (count each publisher once). If all the support comes from one publisher or from vendors only, label the finding "single-perspective".
Do not treat a vendor's own marketing page as evidence of how well its product works. You may cite it for what the vendor says it offers.
Output: a short summary, then a table of findings with columns: finding, supporting sources, publisher type, possible conflict, independent support (yes or no).
Fill in: QUESTION INTERESTED_PARTIES
Especially useful for product, tool or policy comparisons.
Best for: Mixing private and public sources safely
I have supplied my own material and also want you to research the wider web. Keep the two clearly apart.
Question: {{QUESTION}}
What my material contains (one line per file or source): {{MY_MATERIAL}}
What I want the web to add: {{WHAT_WEB_SHOULD_ADD}}
In the report, label every claim with its origin:
- [MINE] when it comes from my supplied material, with the file name and location.
- [WEB] when it comes from a public source, with name, link and date.
- [BOTH] when the two agree, and [CONFLICT] when they disagree, showing both versions side by side.
Do not blend the two without a label. Do not change or "correct" my material to match the web. If my material appears out of date, say so in a separate note with the evidence.
End with a short list: what the web confirmed, what it contradicted, and what my material covers that the web does not.
Fill in: QUESTION MY_MATERIAL WHAT_WEB_SHOULD_ADD
Use where your tool can combine uploaded or connected files with web sources. Check your plan and settings; the sources a tool can use differ by product.
Best for: A method and coverage note on every report
Add this appendix requirement to my deep research brief, so the report shows how the run searched and what it could not reach.
Question: {{QUESTION}}
Source types I most want checked: {{PRIORITY_SOURCE_TYPES}}
End the report with a section "Method and access" that states:
1. The kinds of source you searched or read, which of my priority source types you did not manage to use, and what you used instead.
2. The date range your sources cover, with the oldest and newest source dates.
3. Search angles that returned nothing useful, and topics where you found little or nothing.
4. Sources you found but could not read in full (paywalls, login walls, broken pages, scanned documents), with the link.
5. Every claim that rests only on a snippet, abstract or summary rather than the full document. Mark the same claims "summary only" in the body of the report.
6. Sources you opened but set aside, and why.
Do not imply you read a document you did not. If you cannot say something about your own search, write "not available to me" instead of guessing. Keep the section factual and short.
Fill in: QUESTION PRIORITY_SOURCE_TYPES
Add to the end of any brief. The tool's account of its own search may be incomplete, so treat it as a list of leads and pair it with the source-list audit.
Ask for a Report You Can Audit
The format decides how easily you can check the result. These prompts ask for an answer-first report that exports cleanly, comparison tables where every cell carries a source or says 'not found', market fact tables and dated timelines instead of one smooth narrative.
Best for: A report a busy reader can act on
Write the output-format block for my deep research brief. The report will go to a reader who needs to act on it, and I may export it to a document, so it must read well without editing.
Reader: {{READER}}
Decision they need to make: {{DECISION}}
Maximum length: {{MAX_LENGTH}}
The block should tell the tool to use this structure:
1. Bottom line: the answer in no more than 5 sentences, with a confidence level (high, medium or low) and the reason for it.
2. Key findings: 5 to 8 numbered sections, each heading stating its finding.
3. Where sources disagree, and which side has stronger evidence.
4. Implications for the decision, as options rather than instructions.
5. Limits and open questions, keeping every warning about weak or missing evidence.
6. Sources: one numbered list with title, publisher, date and link, matching numbered markers in the text.
It should also ask for tables wherever more than three comparable figures appear, and no filler introduction. If the report would run over the length, the tool cuts detail rather than caveats and lists at the end what it cut.
Output only the block, written as instructions to the tool, under 150 words.
Fill in: READER DECISION MAX_LENGTH
Paste the block in as the PERFORMANCE part of the one-shot brief. OpenAI's developer guide says to include the expected output format in the prompt. Google's help page says a Gemini report can be exported with Export to Docs, under Share & export in the Canvas panel.
Best for: Option comparisons you can check
Deep research brief: compare the options below and give me a table I can check. Research each option against every criterion, not only the criteria its own marketing covers.
Options to compare: {{OPTIONS}}
Criteria that matter to me, in priority order: {{CRITERIA}}
My situation, in a sentence: {{MY_SITUATION}}
Produce:
1. A comparison table: one row per criterion, one column per option. Every cell holds either a short finding with a source number in brackets, or the words "not found" with no source number. Never estimate to fill a cell.
2. Under the table, a numbered source list with name, link and date.
3. A short section "Where the data is weak" naming cells that rest on a vendor's own claims, on undated sources or on a single source.
4. A recommendation for my situation, clearly separated from the facts, explaining which criteria drove it.
Do not rank options by overall score unless you show the weights. Do not carry assumptions from one row to another. If an option has changed recently, give the date of the change.
Fill in: OPTIONS CRITERIA MY_SITUATION
The classic comparison report, but with an audit trail in every cell.
Best for: Sourced market facts
Run deep research on the market below and return facts in a table format that I can verify, not a story.
Market or segment: {{MARKET}}
Region: {{REGION}}
Period: {{PERIOD}}
What I need to decide: {{DECISION}}
Return a table with these columns: topic, fact, figure or detail, source, publication date, and whether the source is independent or from an interested party. Group rows under: size and growth, customers and segments, main players, pricing, regulation, recent changes.
Rules:
- Every row needs a source and a date. Do not include a row without both.
- If different sources give different figures, give each in its own row and say why they might differ (definition, year, method).
- Do not average or choose between conflicting figures unless a source explains the difference.
- After the table, list the five most important gaps, as questions I could answer by phone calls or surveys.
Fill in: MARKET REGION PERIOD DECISION
A fact table is easier to check than prose, and works with the verification prompts later in the pack.
Best for: Chronologies you can verify
Run deep research and build a timeline.
Subject: {{SUBJECT}}
Start and end of the period: {{PERIOD}}
Type of events to include: {{EVENT_TYPES}}
Return:
1. A table: date, event, source, and a note on whether the date was confirmed by more than one source.
2. Events in strict date order. If a date is approximate or sources disagree, write the range and say which sources give which date.
3. A short list "Cause and effect claimed by sources" showing links between events that a source asserts, with the source named. Do not infer causes yourself.
4. A short list of gaps in the timeline.
Do not add events from memory. If you cannot find a source for an event you believe happened, list it under "Unsourced" and do not put it in the table. Keep commentary minimal; the value is the table.
Fill in: SUBJECT PERIOD EVENT_TYPES
Dates are easy to check against sources, which makes a timeline a useful format for a first look at a tool's accuracy.
The prompts in this section rely on features described on the vendors' own help pages: Gemini's editable research plan and custom visualisations, Claude's research across connected tools and a record of which ones it used, and Perplexity Computer's same-session follow-on. The note under each prompt says what the page says. Plans and settings change, so check the current page.
Best for: Editing the plan before the run
Gemini has shown me a research plan for my topic. Review it before I click Start research.
My research question and brief: {{MY_BRIEF}}
The plan as Gemini wrote it, pasted: {{PROPOSED_PLAN}}
Check the plan against my brief and tell me:
1. Which parts of my brief the plan covers and which it narrows or skips.
2. Steps likely to use source types I said to avoid, or that ignore my source rules.
3. Steps that duplicate each other or look unnecessary.
4. Steps to add, written as short instructions I can type into Edit plan.
5. Whether the plan covers the date range, conflicting sources and the output format I asked for.
Finish with a revised numbered plan I can paste over the original, keeping Gemini's wording where it was fine.
Fill in: MY_BRIEF PROPOSED_PLAN
Google's help page says Gemini creates a research plan, that you click Edit plan to update it before you create a report, and then click Start research. Run this in a separate chat while the plan waits.
Best for: Charts from report data only
Using only the figures in the research report above, create a visualisation.
What I want to show: {{WHAT_TO_SHOW}}
Chart or diagram type (or "you choose"): {{CHART_TYPE}}
Who will see it: {{AUDIENCE}}
Rules:
- Use only numbers that appear in the report. Do not add data, estimates or trend lines.
- Label axes and units, and give the source and date of each figure in a caption.
- If figures come from sources with different definitions or years, say so in a note on the chart.
- If the report lacks the data for what I asked, tell me what is missing.
After the visual, list the figures used with their source numbers from the report.
Fill in: WHAT_TO_SHOW CHART_TYPE AUDIENCE
Google's help page says: 'To create a custom visualization based on the report, enter the description as a prompt in the text box.' It also says this isn't available if you include Gmail or Drive as a source.
Best for: Research that uses connectors within limits you set
Please use the research tool for the question below. You can reach my connected tools during research, so here are the limits I want you to keep to.
Question: {{QUESTION}}
Connected sources to search, and what to look for in each: {{SOURCES_TO_SEARCH}}
Connected sources, folders or subjects to leave alone: {{OFF_LIMITS}}
What the web should add: {{WEB_ASSIGNMENT}}
Rules for this run:
- Start with the connected sources I named, so you know what we already have, then research the web for what is missing.
- Search only the connected sources I named, for the items described. If another connected source looks relevant, mention it in the report instead of searching it.
- Quote personal details from emails or calendar entries only where the question needs them.
- Label each finding [CONNECTED: source name] or [WEB: link and date].
- End with a list of each connected source you searched and what you searched it for.
If keeping to these limits leaves part of the question unanswered, say which part rather than widening the search.
Fill in: QUESTION SOURCES_TO_SEARCH OFF_LIMITS WEB_ASSIGNMENT
Anthropic's help page: Research needs a paid plan and web search turned on, and starts from the + menu; Claude researches the web and connected tools such as Gmail, Google Calendar and Google Docs, and can use connector tools without asking approval each time. A prompt is a request, not a setting: disconnect anything Claude must not reach.
Best for: Confirming connectors were used
Before I rely on this research, tell me exactly which of my connected sources contributed to it.
Connected tools I expected you to use: {{EXPECTED_CONNECTORS}}
The research question: {{QUESTION}}
For each connected tool I listed, say:
- Whether you retrieved anything from it for this research. Answer yes, no, or not sure.
- If yes, which items (titles or subjects) and which findings in the report depend on them.
- If no, what you can tell from this conversation about why. Say plainly when you cannot tell.
Then list findings in the report that came only from the web. If a tool I expected returned nothing, give me a one-line prompt to run the search again with that source named.
Fill in: EXPECTED_CONNECTORS QUESTION
Anthropic's help page says Research in Claude mobile only searches the web and does not use your connectors, so run research on web or desktop. The answer here is Claude's own account; check it against the connector's own history before relying on it.
Best for: A slide outline that keeps its sources
Continue in this same session. Using the research report you just produced, create a slide deck outline.
Audience: {{AUDIENCE}}
Number of content slides (not counting the final open-questions slide): {{SLIDE_COUNT}}
The one message I want them to remember: {{KEY_MESSAGE}}
For each slide give: a title that states the finding, 2 to 4 bullet points using only facts from the report, the chart or table to show (if any), and the source name and date for every figure on the slide. Add a final slide listing open questions and weak evidence exactly as the report described them.
Do not add claims or figures that are not in the report. If the report cannot support a slide I asked for, say so and drop the slide.
Fill in: AUDIENCE SLIDE_COUNT KEY_MESSAGE
Perplexity's help page for Computer says you can continue in the same session, for example by asking Computer to turn a report into a slide deck. Computer is one Perplexity product, not available to every account, so check your plan.
Check What the Tool Returned
These prompts act on the tool's own output: its claims, its cited source list, its numbers and its source dates. They also test the tool against a question where you know the answer, compare it with a plain answer, compare two tools' reports on the same brief, and re-check chosen claims in an independent run.
Best for: Seeing what the report really leans on
Audit the source list of a research report. I am pasting the list as the tool gave it, with the number of claims in the report that cite each source.
Source list with claim counts: {{SOURCE_LIST}}
Produce:
1. A table: source, publisher, publisher type (independent research body, regulator, trade body, vendor or seller, news outlet, blog, unknown), date, and number of claims that cite it. Use "undated" where there is no date.
2. Publishers that appear more than once under different titles. Count them together.
3. Which publisher accounts for the most citations, and how many of the total citations come from vendors or sellers.
4. Sources that are likely to be summaries of other sources rather than original.
5. The three sources I should open first, with the reason for each.
Work only from the list. Do not judge whether a source is accurate; judge only what kind it is and how much of the report leans on it.
Fill in: SOURCE_LIST
Paste the tool's source list, adding the claim counts yourself. Shows at a glance when a report leans on one publisher.
Example output · Claude (example written by The Prompt Index)
**Input (SOURCE_LIST, with claim counts added by me):**
1. Northern Practice Association, "Software Survey 2024", full report. Cited by 3 claims.
2. Vendor Y, pricing page, accessed October 2026. Cited by 2 claims.
3. Vendor Y blog, "Why firms switch to cloud", undated. Cited by 3 claims.
4. Accountancy blog, "Ten reasons to go cloud", 2022. Cited by 1 claim.
5. Northern Practice Association, "Survey 2024: key findings", summary page. Cited by 1 claim.
**Output:**
| Source | Publisher | Publisher type | Date | Claims citing it |
|---|---|---|---|---|
| 1. Software Survey 2024 | Northern Practice Association | trade body (inferred from the name) | 2024 | 3 |
| 2. Pricing page | Vendor Y | vendor or seller | October 2026 (accessed) | 2 |
| 3. "Why firms switch to cloud" | Vendor Y | vendor or seller | undated | 3 |
| 4. "Ten reasons to go cloud" | unknown | blog | 2022 | 1 |
| 5. "Survey 2024: key findings" | Northern Practice Association | trade body (inferred from the name) | not shown | 1 |
**Same publisher under different titles:** sources 1 and 5 (Northern Practice Association, 3 + 1 = 4 citations); sources 2 and 3 (Vendor Y, 2 + 3 = 5 citations).
**Most citations:** Vendor Y, 5 of the 10 total citations (3 + 2 + 3 + 1 + 1 = 10). All 5 come from a vendor or seller.
**Likely summaries:** source 5 appears to be a summary of source 1, going by the titles.
**Open first:**
1. Source 3: undated, from a vendor, and cited by 3 claims.
2. Source 1: the full survey; compare it with the summary in source 5.
3. Source 4: a 2022 blog, older than the other dated sources, with an unknown publisher.
Best for: Seeing which claims lean on which source
Map the citations in a research report. The report below came from an AI research tool and carries its own source numbers.
Report text with its source markers: {{REPORT_TEXT}}
The report's numbered source list: {{REPORT_SOURCE_LIST}}
Make a table with one row per distinct factual claim (figure, date, quote, named study, cause-and-effect or comparison), in order. Columns:
1. Number.
2. Claim, 25 words or fewer, faithful to the report.
3. Source numbers the report attaches to it, or NO SOURCE.
4. Support type: DIRECT (a source number sits on this claim), DERIVED (calculated from other claims in the report; name them), or UNATTACHED (a source is cited somewhere nearby, but not on this claim).
5. Wording flag: STRONG WORDING if it says "always", "proves", "the best" or similar.
6. What to open to check it.
Do not judge whether any claim is true. Finish with counts of each support type and a list of source numbers that carry more than three claims.
Fill in: REPORT_TEXT REPORT_SOURCE_LIST
Paste the whole report with its source markers, plus its source list. Shows which claims actually have a source on them, which is often fewer than the footnote count suggests.
Example output · Claude (example written by The Prompt Index)
**Input (REPORT_TEXT):**
> Cloud practice-management software always reduces admin time for small accountancy firms. A 2024 survey by the Regional Accountants Institute found that 61% of responding firms, 61 of 100, reported time savings [1]. Vendor X lists its price as 30 pounds per user per month [2], so a 12-user firm would pay 4,320 pounds a year and 12,960 pounds over three years. Firms of 12 users save about 15 hours a week, which is 780 hours a year, worth 9,000 pounds a year at 20 pounds an hour. Migration usually takes 6 to 8 weeks.
**Input (REPORT_SOURCE_LIST):** [1] Regional Accountants Institute, Practice Software Survey 2024. [2] Vendor X pricing page, accessed October 2026.
**Output:**
| # | Claim | Source numbers | Support type | Wording flag | What to open |
|---|---|---|---|---|---|
| 1 | Cloud practice-management software always reduces admin time for small accountancy firms. | NO SOURCE | UNATTACHED (sources [1] and [2] appear nearby, none on this claim) | STRONG WORDING | The study that measured admin time, not a vendor page. |
| 2 | 61% of responding firms (61 of 100) reported time savings in a 2024 survey. | [1] | DIRECT | none | Source [1]: question wording, who responded. |
| 3 | Vendor X lists its price as 30 pounds per user per month. | [2] | DIRECT | none | Vendor X's current pricing page. |
| 4 | A 12-user firm would pay 4,320 pounds a year. | NO SOURCE | DERIVED from claims 3 and the 12-user figure | none | Recompute; check billing terms. |
| 5 | The cost over three years would be 12,960 pounds. | NO SOURCE | DERIVED from claim 4 | none | Recompute. |
| 6 | Firms of 12 users save about 15 hours a week. | NO SOURCE | UNATTACHED | none | The data behind the estimate. |
| 7 | That is 780 hours a year. | NO SOURCE | DERIVED from claim 6 | none | Recompute. |
| 8 | Those hours are worth 9,000 pounds a year at 20 pounds an hour. | NO SOURCE | DERIVED from claim 7 and the hourly rate (which has no source) | none | Recompute; find the source of 20 pounds an hour. |
| 9 | Migration usually takes 6 to 8 weeks. | NO SOURCE | UNATTACHED | none | Vendor implementation guides and accounts from similar firms. |
Counts: DIRECT 2, DERIVED 4, UNATTACHED 3 (total 9). STRONG WORDING: 1. Source numbers carrying more than three claims: none ([1] and [2] each carry one).
Best for: Testing a tool before trusting it
Help me test how reliable a deep research tool is on a subject where I already know the answers.
A subject I know well and can check: {{SUBJECT_I_KNOW}}
Facts I know for certain about it: {{FACTS_I_KNOW}}
How I know them (my own records, direct experience, an official document): {{HOW_I_KNOW}}
Create:
1. A research question about the subject that a good report should be able to answer, which is not too easy and not too obscure. It should touch at least 3 of my known facts.
2. The brief to paste into the research tool, in two parts: the question and a format request that puts each claim next to a source.
3. An answer key: a numbered list of my known facts, which I will use for scoring.
4. Two "trap" items: details a careless report could get wrong, such as an old figure that has since changed or a commonly repeated misstatement. Only include traps I can confirm from what I gave you.
Do not add facts I did not give. If I gave too few to build a fair test, say what I need to add.
Fill in: SUBJECT_I_KNOW FACTS_I_KNOW HOW_I_KNOW
Zapier's guide suggests testing the tool on something where you already have a solid expectation. This sets the test up properly.
Best for: Scoring the tool honestly
Score the research tool's output against my answer key. Use only what I provide.
Research question: {{QUESTION}}
My answer key (numbered facts I know to be true): {{ANSWER_KEY}}
Excerpt from the tool's report: {{REPORT_EXCERPT}}
For each numbered fact in my key, say whether the report: states it correctly, states it partly (what is missing), contradicts it (quote the report), or does not mention it.
Then:
- List claims in the excerpt that are not in my key, so I can check them separately. Do not mark them right or wrong.
- Note whether each claim you scored has a source attached in the excerpt.
- Give a tally: correct, partial, contradicted, missing.
- Say in two sentences what the tally implies about how much checking other parts of the report will need, and what it cannot tell me.
Fill in: QUESTION ANSWER_KEY REPORT_EXCERPT
Use after a calibration run. A small tally from one excerpt is a prompt to check, not a measure of overall accuracy.
Example output · Claude (example written by The Prompt Index)
**Input**
- QUESTION: What does a PhD student at Northfield University (a fictional university) need to do to submit a thesis for June graduation?
- ANSWER_KEY: 1. Thesis is submitted through the Graduate Office portal. 2. Deadline for June graduation is 15 March. 3. Maximum length is 80,000 words, excluding appendices.
- REPORT_EXCERPT: "Students submit their thesis electronically via the Graduate Office portal [1]. The submission deadline for June graduation is 1 April [2]. Theses are typically between 80,000 and 100,000 words [3]. A hard-bound copy must also be deposited in the library [4]."
**Output**
| Key fact | Verdict | Evidence |
|---|---|---|
| 1. Portal submission | States it correctly | "submit their thesis electronically via the Graduate Office portal [1]" |
| 2. Deadline 15 March | Contradicts it | Report says "1 April [2]" |
| 3. Maximum 80,000 words, excluding appendices | Contradicts it | Report says "typically between 80,000 and 100,000 words [3]"; a range that runs to 100,000 conflicts with a maximum of 80,000, and it does not mention the appendix rule |
**Claims in the excerpt that are not in my key (check separately):** hard-bound copy deposited in the library [4].
**Sources attached:** all four statements carry a source number; I cannot see what [1] to [4] are.
**Tally:** correct 1, partial 0, contradicted 2, missing 0.
**What it implies:** two of three key facts are contradicted, so dates and limits elsewhere in this report need checking against official pages. It cannot tell you how accurate the whole report is, because one excerpt of three facts is a very small sample.
Best for: Judging whether to run it next time
I asked the same question twice: once in a deep research mode and once as an ordinary quick answer. Compare them.
The question: {{QUESTION}}
Deep research report: {{DEEP_REPORT}}
Ordinary quick answer: {{QUICK_ANSWER}}
Tell me:
1. Points that appear in both.
2. Points that appear only in the deep research report, and whether each has a source.
3. Points in the quick answer that the deep research report omits or contradicts.
4. Whether the extra material changes the answer to my question, or only adds detail.
5. Which of the report's sources the quick answer did not use, such as recent pages or documents several links deep. If the quick answer cites no sources, say so and skip this point rather than guessing what it relied on.
6. Anything in either that needs checking before I use it.
Do not say which is "better" in general. Say only what the deep research run added or lost for this question.
Fill in: QUESTION DEEP_REPORT QUICK_ANSWER
A cheap way to learn when deep research earns its time for your type of question.
Best for: Using disagreement to find errors
I gave the same research brief to two different tools. Compare what came back.
The brief: {{ORIGINAL_BRIEF}}
Report from tool A: {{REPORT_A}}
Report from tool B: {{REPORT_B}}
Produce:
1. Findings both reports share, with the sources each cites.
2. Findings in only one report.
3. Direct contradictions: claim A versus claim B, with each side's source.
4. Differences in coverage: topics or source types one handled and the other did not.
5. The five items I should check against primary sources, chosen because the reports conflict or only one mentions them.
Do not resolve a contradiction yourself. For each, describe the sources behind each side (primary or secondary, date, publisher type) and leave the decision to me.
Fill in: ORIGINAL_BRIEF REPORT_A REPORT_B
Disagreement between tools shows you where to look first. If the two reports exceed your chat's limit, compare them section by section.
Best for: Arithmetic and unit checks
Check the arithmetic and consistency of every number in the deep research report below. Reports like this combine figures from many sources, so a total can mix units, years or definitions. Do not use outside facts.
Report: {{REPORT_TEXT}}
Do the following:
1. List every figure with its unit, period and what it measures.
2. Where one figure can be derived from others (totals, percentages, averages, annual versus monthly, per-person versus total), recompute it step by step and show the working. Mark each one "matches", "does not match" (give the correct value) or "cannot check".
3. Flag inconsistencies: the same quantity given with different values, mixed currencies or units, mixed time periods, and percentages that do not add up.
4. Flag figures that sound precise but have no stated method or source.
Finish with a list of the errors found, ranked by how much they would change a decision. If everything checks out, say so and say what I still cannot verify without the original sources.
Fill in: REPORT_TEXT
A chat model can also slip on arithmetic, so recheck any mismatch it reports with a calculator before acting on it.
Example output · Claude (example written by The Prompt Index)
**Input (REPORT_TEXT):**
> Cloud practice-management software always reduces admin time for small accountancy firms. A 2024 survey by the Regional Accountants Institute found that 61% of responding firms, 61 of 100, reported time savings [1]. Vendor X lists its price as 30 pounds per user per month [2], so a 12-user firm would pay 4,320 pounds a year and 12,960 pounds over three years. Firms of 12 users save about 15 hours a week, which is 780 hours a year, worth 9,000 pounds a year at 20 pounds an hour. Migration usually takes 6 to 8 weeks.
**Output:**
**1. Figures:** 61% and 61 of 100 (survey); 30 pounds per user per month; 12 users; 4,320 pounds a year; 12,960 pounds over three years; 15 hours a week; 780 hours a year; 9,000 pounds a year; 20 pounds an hour; 6 to 8 weeks.
**2. Recomputation**
- 61 of 100 = 61/100 = 61%: **matches**.
- 12 users x 30 pounds = 360 pounds a month; 360 x 12 months = 4,320 pounds: **matches**.
- 4,320 x 3 years = 12,960 pounds: **matches** (assumes the price does not change).
- 15 hours x 52 weeks = 780 hours: **matches** (assumes 52 working weeks).
- 780 hours x 20 pounds = 15,600 pounds. The report says 9,000: **does not match**. Correct value from the report's own inputs: 15,600; the difference is 15,600 - 9,000 = 6,600 pounds. The report's 9,000 would correspond to 9,000 / 20 = 450 hours.
- 6 to 8 weeks: **cannot check**; nothing to derive it from.
**3. Inconsistencies:** only the value of the saved hours. All money is in pounds; the monthly price is converted to annual in the working above.
**4. Precise-sounding figures with no method or source:** 15 hours a week, 20 pounds an hour, 6 to 8 weeks.
**Errors ranked by effect on a decision:**
1. Value of saved time: 9,000 stated, 15,600 from the report's own hours and rate. Either the hours or the total is wrong, and the report does not say which.
2. The 15-hour estimate has no source, so the corrected total is also unverified.
Best for: Catching expired facts
Check the deep research report below for information that may be out of date. The run used sources of many ages, and the report may present them all in the present tense.
Report, with its source list: {{REPORT_TEXT}}
Date the run was done: {{RUN_DATE}}
Date I plan to act on it: {{ACTION_DATE}}
List every item that could have changed since its source was written: prices, plans and features, rules and regulations, rankings, headcounts, product names and versions, office holders, and statistics that are updated on a schedule.
For each item give:
- The quoted text and its source date from the report's source list, or "undated".
- The gap between the source date and the date I plan to act.
- How quickly this kind of information usually changes (fast, medium, slow).
- Where I should check the current value (official pricing page, regulator, statistical release).
Rank the list by how much the report depends on the item. Do not supply updated values from memory; you may be out of date yourself, and that is the point of this check.
Fill in: REPORT_TEXT RUN_DATE ACTION_DATE
Reports can be accurate on the day yet stale by the time you act.
Best for: Time-boxed fact-checking
Help me plan a time-boxed check of a deep research report before I rely on it.
Decision the report will inform: {{DECISION}}
The report's conclusion: {{CONCLUSION}}
The report's key claims, with their source numbers: {{KEY_CLAIMS}}
Its method and access notes (sources it could not read, claims resting on summaries), or "none given": {{ACCESS_NOTES}}
Time I have for checking: {{TIME_AVAILABLE}}
Build a checking plan:
1. Work out which claims the conclusion depends on, and rank them by how much a mistake in each would change my decision. Move up any claim that rests on a summary, an abstract or a source the run could not read in full.
2. For the top claims that fit my time, say exactly what to open (the regulator's page, the company's filing, the original paper, the vendor's pricing page), what to look for, and what result would mean the claim is wrong.
3. A tick-box list I can work through in order.
4. Claims I will not have time to check, and a plain wording that marks them "unchecked" in whatever I write next.
Do not say whether any claim is true or false. Keep the plan realistic for the time available.
Fill in: DECISION CONCLUSION KEY_CLAIMS ACCESS_NOTES TIME_AVAILABLE
Turns 'I should verify this' into a short, ordered list.
Best for: A second, independent check
Independently verify the claims below. You have not seen the report they came from.
Claims to verify: {{CLAIMS_LIST}}
Region: {{REGION}}
Period: {{PERIOD}}
For each claim, search for evidence yourself and report:
- Verdict: supported, partly supported, contradicted, or unable to verify.
- The best sources, with links and dates, preferring primary sources.
- Any difference between the claim and the source, such as a different year, scope or number.
- If you find a conflicting figure, give it with its source.
Do not assume a claim is right because it is plausible. Present the results in a table, then list the claims where your findings differ from what was asserted.
Fill in: CLAIMS_LIST REGION PERIOD
Run it in a different tool from the one that wrote the report, or in a fresh session. Paste only the claims, not the report.
Re-Run Without Starting Over
Most useful reports need a second run. These prompts turn gaps into a narrower brief, go deeper on the weakest section, re-run with stricter source rules after the first report broke them, and refresh an older report with only what changed.
Best for: A focused follow-up run
Here is a research report and the brief it answered. Write the brief for a second, narrower run that fills its gaps.
Original brief: {{ORIGINAL_BRIEF}}
Report: {{REPORT_TEXT}}
What I already know that the report did not mention: {{WHAT_I_KNOW}}
Do the following:
1. List the gaps: questions in my brief that were not answered, claims that rested on weak sources, and "not found" items.
2. Given the decision this feeds, which is {{DECISION}}, choose the three gaps where closing them could change the answer.
3. Write a new brief that targets those three only. Include what we already established, from the report and my notes (so it is not repeated), the exact questions, tighter source rules and the output format.
4. Add one instruction: where the new findings contradict the first report, say so and show both sources.
Keep the new brief under 300 words.
Fill in: ORIGINAL_BRIEF REPORT_TEXT WHAT_I_KNOW DECISION
Second runs work best when they are narrower than the first.
Best for: Targeted improvement
Run a new deep research pass on one section of an earlier report. I am pasting only that section, so stay within it.
Section: {{SECTION_TEXT}}
Why I think it is weak: {{WHY_WEAK}}
Sources I trust for this topic: {{TRUSTED_SOURCES}}
Research specifically to improve it. Report:
1. What the original section got right, wrong or left out, citing sources.
2. A rewritten section using only claims you can source, each with a link and date.
3. Any claim from the original section that you could not support, marked "unsupported" rather than silently dropped.
4. Anything you found that changes the conclusion of the original report.
Do not widen the scope beyond this section. If you find nothing better than the original, say so.
Fill in: SECTION_TEXT WHY_WEAK TRUSTED_SOURCES
Stronger than a vague 'make it better'.
Best for: Enforcing your rules on a second run
My first research run did not follow all my source rules. Write a stricter brief for a re-run.
My original source rules: {{ORIGINAL_SOURCE_RULES}}
The report's source list: {{SOURCE_LIST}}
The question: {{QUESTION}}
Do the following:
1. Compare the source list with my rules and list each breach: a source type I excluded, a missing date, a secondary source where I asked for primary, or reliance on one publisher.
2. For each breach, say which claims in the report I should check, going by where that source's number appears in the report's text if I pasted any.
3. Write a new brief for the re-run with the original question, my rules rewritten more tightly where they were ignored, and a requirement that the report ends with a table showing, for each source, which of my rules it satisfies.
Keep the brief under 300 words.
Fill in: ORIGINAL_SOURCE_RULES SOURCE_LIST QUESTION
Source rules are often only partly followed. This checks the list and tightens the brief.
Best for: Updating without starting over
Update an earlier piece of research. I want only what has changed.
Topic: {{TOPIC}}
Date of my previous report: {{PREVIOUS_DATE}}
The key findings of that report: {{PREVIOUS_FINDINGS}}
Research the topic again, focusing on developments after the date above. Report:
1. Findings that are unchanged, briefly, with a current source and date.
2. Findings that have changed, with the old statement, the new one, and the source and date.
3. New developments that were not in my earlier findings.
4. Anything in my earlier findings you could not re-confirm.
Do not rewrite the whole topic from scratch. Make clear which items come from sources dated after my previous report. If you cannot find dated sources after that date, say so.
Fill in: TOPIC PREVIOUS_DATE PREVIOUS_FINDINGS
For recurring research such as quarterly market or regulatory checks.
From Report to Decision
These prompts map the report's findings onto your decision criteria with evidence tags, choose between acting, checking and re-running, and record what the run did and did not establish.
Best for: Seeing which cells are evidence and which are gaps
I have a research report and a decision to make. Map what the report says onto my criteria without adding anything.
Criteria: {{CRITERIA}}
Options: {{OPTIONS}}
The report's findings, each with its source number as the report gave it: {{FINDINGS_WITH_SOURCES}}
Build a table: options in rows, criteria in columns. Each cell holds the relevant finding in a few words, its source number, and one of three tags: SUPPORTED (a source states it directly), INFERRED (the finding implies it but does not state it), or NO EVIDENCE. Leave a cell as NO EVIDENCE rather than filling it.
Below the table list: criteria where the report had nothing for any option, options with the most NO EVIDENCE cells, and findings the report gave that fit none of my criteria, which may mean I am missing a criterion. Do not score or rank the options.
Fill in: CRITERIA OPTIONS FINDINGS_WITH_SOURCES
The tags keep the table honest about what the research actually established.
Best for: An audit trail for research-based decisions
Write a research log entry for a deep research run so someone else can repeat or challenge it later.
Date of the run: {{RUN_DATE}}
Tool and mode used: {{TOOL_AND_MODE}}
The brief I gave (pasted): {{BRIEF}}
The research plan or method note the tool showed, or "none": {{PLAN_OR_METHOD}}
The report's conclusion and caveats (pasted): {{REPORT_SUMMARY}}
What I checked myself, and what I did not: {{MY_CHECKS}}
Write the entry with these headings: Question, Brief (verbatim), Tool and date, Plan or method as the tool described it, Conclusion as the tool stated it, Caveats the tool gave, What I verified (with how), What I did not verify, Known weaknesses, and Next steps.
Use only what I provided. Do not claim a check was done unless I said so. Keep the entry under 250 words.
Fill in: RUN_DATE TOOL_AND_MODE BRIEF PLAN_OR_METHOD REPORT_SUMMARY MY_CHECKS
A short record you can attach to the decision that relied on the report.
Best for: Choosing the next step after a run
A deep research run has finished. Help me decide what to do with the report next: act on it, check specific claims first, or write a narrower second brief.
Decision the report feeds: {{DECISION}}
The report's bottom line and caveats: {{REPORT_SUMMARY}}
Its method and access notes, or "none given": {{ACCESS_NOTES}}
Results of any checks I have run (citation map, number check, calibration score), or "none yet": {{CHECK_RESULTS}}
Answer each question and show your reasoning:
1. Does the bottom line rest on claims that carry a source number, or on claims with no source or derived from other claims?
2. Did the run reach the source types this decision needs, or do the access notes show gaps?
3. Do my check results show errors in the parts the conclusion depends on?
4. Would acting on a wrong conclusion be cheap to undo or costly?
Then recommend one of: ACT (say what the decision may rely on), CHECK FIRST (name up to five claims and what to open for each), or RE-RUN (name the gap a second brief must close). Do not judge whether any claim is true; work only from what I gave you.
Fill in: DECISION REPORT_SUMMARY ACCESS_NOTES CHECK_RESULTS
Use after the checking prompts. It turns their results into one next step.