A structured prompt library for every stage of research — framing the question, screening sources, extracting evidence, challenging your own conclusion and auditing the final draft.
Free. We email you the file — no account, no card.
+ 61 more rows · 2 tabs · fully editable
What you’ll get
Each one written for a specific job, not a generic "act as a researcher" template.
Every prompt is tagged Define through Deliver, so you can jump to the part you are stuck on.
Beginner, Intermediate or Advanced — 21 beginner prompts to start with, 21 advanced ones to grow into.
Supported models plus the one each prompt was optimised for. Some have GPT, Claude and Gemini variants.
Roughly what each prompt costs before you paste it, from 180 to 850 tokens.
Filter by things like provenance, confirmation bias, evidence extraction or overclaiming.
How the variables work, what the columns mean, and three suggested paths through the pack.
No app, no login, no locked cells. Filter it, rewrite prompts, add your own rows.
Inside the pack
Real excerpts from the workbook, abbreviated. If these look useful, the other 58 follow the same standard.
Turns a rough topic into research questions worth actually investigating.
Excerpt
You are an expert research design assistant. Your job is to help me turn a topic, scope, or rough idea into strong research questions that are clear, useful, and worth investigating. ## Topic {{topic}} ## Current scope {{scope}} ## My goal {{goal}} ## Intended use {{intended_use}} ## Number of primary questions {{primary_question_count}}
GPT · Claude · Gemini
Converts a goal into staged work with priorities and next actions.
Excerpt
You are an expert research planning assistant. Your job is to turn a topic, scope, and research goal into a practical research plan with clear stages, priorities, and next actions. ## Research goal {{goal}} ## Why this research is being done {{reason}} ## Final audience for the eventual output {{final_audience}}
GPT · Claude · Gemini
Decides what evidence would settle each question — before you go looking.
Excerpt
Act as an evidence-planning specialist. Your job is to determine **what evidence must be found before research begins**, so conclusions are driven by evidence rather than by whatever information happens to appear during searching. Create an evidence requirements map connecting each important research question or proposed conclusion to the evidence needed to support, reject, qualify, or leave it unresolved.
GPT · Claude · Gemini
A fast first pass that sorts a source list before you invest reading time.
Excerpt
You are an expert source reliability screening assistant. Your job is to perform an early reliability screen on a set of sources so I can quickly see which ones look stronger, weaker, riskier, narrower, outdated, or more useful before deeper verification begins. ## Sources to screen {{sources}}
Separate GPT, Claude & Gemini versions
Asks whether a study actually supports the claim you want to hang on it.
Excerpt
Act as a critical research-methodology reviewer. Evaluate whether the methodology of the provided study supports the conclusions being drawn from it. Do not judge a study merely by whether its findings seem plausible or agree with expectations. If important methodological information is unavailable, explicitly mark it as **Not Reported / Cannot Assess** rather than guessing.
GPT · Claude · Gemini
Finds out whether ten sources are real corroboration or one source repeated.
Excerpt
Act as a claim-provenance investigator. Trace a repeated claim through the available sources to determine **where it originated, which sources independently support it, and whether apparent corroboration is actually multiple publications repeating the same underlying evidence**. Do not treat repetition as independent verification.
GPT · Claude · Gemini
Checks whether your process ever gave the opposing evidence a fair chance.
Excerpt
Act as an independent research-process auditor. Examine completed or in-progress research for signs that the process may have **favored an initial belief, hypothesis, preferred conclusion, or expected outcome**. Your goal is to determine whether the research process gave credible competing evidence a fair opportunity to affect the result.
GPT · Claude · Gemini
The last gate before a draft is sent, submitted or published.
Excerpt
You are a final trust audit reviewer. Perform a last-pass trust review of the draft before delivery, publication, submission, or client use. Audit across: claim support, citations and attribution, numbers, dates and versions, internal consistency, reasoning integrity, overclaim risk, and source faithfulness.
GPT · Claude · Gemini
The research process
Most bad research fails at a specific step. The Stage column tells you which prompts belong to the step you’re standing on.
Turn a vague topic into a precise, answerable question.
Decide what evidence you need before you start collecting it.
Build search queries and find the right kinds of sources.
Screen sources for reliability, method quality and conflicts.
Pull out claims, quotes and data with their caveats intact.
Put sources side by side and diagnose where they disagree.
Test arguments, assumptions and causal claims.
Turn scattered findings into a position you can defend.
Attack your own conclusion before anyone else does.
Check that every claim traces back to a real source.
Write it up, then run a final trust audit before it ships.
How to use it
Every prompt is plain markdown in a single cell, with {{variables}} marking the parts you replace. If you don’t have a value, say so — don’t let the model invent one.
Filter the Stage column. Framing a question and stress-testing a conclusion need different prompts.
Each prompt lives in one cell as plain markdown. Copy the whole thing, formatting included.
Anything in double braces is yours to replace. If a variable is marked optional and you do not have it, delete the line.
Where a prompt asks for sources, notes or a draft, paste them in or attach the files before you run it.
Feed the output of one prompt into the next. Most work needs three or four prompts, not sixty-six.
Once you have the pack
Workbench is our browser extension for prompt libraries: import a spreadsheet in one go, keep prompts organised by stage and tag, fill the variables in a form instead of by hand, and run them without leaving ChatGPT, Claude or Gemini.
Who it’s for
Methodology reviews, provenance tracing and bias audits.
A structure for literature work that does not start with a blank page.
Client-ready findings with the caveats still attached.
Market and competitor research you can actually bet on.
Claim checking, source comparison and overclaim detection.
Evidence behind the positioning, not just the confident-sounding version.
Example workflow
You don’t need to run all 66 prompts. Every prompt works on its own — a quick piece of research might use three, a rigorous one might use ten. And nothing here replaces checking the important claims against real sources yourself.
Get the pack
Enter your email and we’ll send ai-research-prompt-pack.xlsx straight to your inbox.
We may email you occasionally for feedback, or to tell you about new Cognardo resources. Unsubscribe any time.
FAQ
Yes. Enter your email and we send the file. There is no payment step, no trial, and no card required.
Yes — the pack is written primarily for GPT, and every prompt except the model-specific variants works there as-is.
Yes. The Supported Models column lists which models each prompt is designed for, and several prompts — the search query builders, the quote extractors and the reliability pre-screeners — ship in separate GPT, Claude and Gemini versions.
A single .xlsx workbook with two tabs: Start Here and AI Research Prompts. It opens in Excel, Google Sheets, Numbers and LibreOffice.
No, and you should not. Every prompt works on its own. A quick piece of research might use three. A rigorous one might use ten.
Yes. Nothing is locked. Rewrite them for your field, tighten them, or add your own rows — it is your copy.
Anything inside double braces is a placeholder you replace with your own information — your question, your sources, your audience. If a variable is marked optional and you do not have that information, tell the model it is unavailable rather than inventing it.
No. AI-generated research is not verified evidence. Several prompts in the pack exist specifically to help you check claims against real sources — but for anything consequential, confirm the important claims yourself.