Template · Copy-paste · 7 min read

A validation prompt that argues back

Ask a chatbot whether your idea is good and it will tell you it is. Not because the idea is good, but because you asked a question with an agreeable answer.

The prompts below remove that path: the model has to argue the case against your idea first, name the competitors you forgot, mark which numbers it is guessing, and finish with a verdict it must defend.

Prompt 1: the validation pass

Works in ChatGPT, Claude and Gemini. Turn web search on if you have it, and fill the two blocks in brackets.

You are evaluating a business idea. Your job is to be useful, not agreeable. THE IDEA <idea> [paste your idea here — what it does, for whom, how it makes money] </idea> WHAT I ALREADY KNOW <context> [who you are, what you have built, who you have talked to, what they said, what you are unsure about. If you skip this, say so — do not invent it.] </context> Work through these steps in order. Do not skip ahead to the verdict. 1. STEELMAN THE OPPOSITE Argue, in five sentences, why this idea fails. Not generic risks — the specific reason THIS idea dies. Name the failure mode. 2. WHO ALREADY DOES THIS List real competitors and near-substitutes, including the boring ones (a spreadsheet, an agency, doing nothing). For each: what they charge and why someone would leave them. If you are unsure a company exists, mark it UNVERIFIED rather than dropping it. 3. THE CUSTOMER'S ACTUAL ALTERNATIVE What does this person do today, at the exact moment the problem shows up? How much does that cost them in money, time or embarrassment? If the honest answer is "they live with it", say that. 4. WHAT WOULD HAVE TO BE TRUE List the assumptions this idea rests on, ordered by how badly it dies if each is wrong. For each: how to test it this week for under $100. 5. NUMBERS, WITH CONFIDENCE MARKED Market size, willingness to pay, cost to serve. Mark every figure as KNOWN (with a source), ESTIMATED (with the reasoning), or GUESSED. Do not produce a figure you would not defend. 6. VERDICT Score each 1-5: problem severity, evidence it exists, size of prize, ability to reach these people, defensibility, founder fit. Then: GO, PIVOT or NO-GO, with the single sentence that decides it. State what evidence would flip your verdict. Rules: no praise, no encouragement, no summary of what I said back to me. If the context is too thin to answer a section, write "not enough input" and say exactly what you would need.

Prompt 2: find the complaints

Run this second. It looks for how people describe the problem in their own words, searching for complaints rather than recommendations.

Find how people describe this problem in their own words. <problem> [the problem your idea solves, in one sentence] </problem> Search for real discussions where people complain about this — forums, communities, review sites. Use search terms that find complaints, not recommendations: pair the topic with phrases like "waste of money", "tired of", "gave up on", "regret", "not working for us". Return 10 quotes. For each: the quote verbatim, where it came from, and what the person was actually trying to do when it went wrong. Then: which three phrases repeat across different people? Those are the words to use on your landing page. If you cannot find real quotes, say so — do not write examples yourself.

Now test the answer you got

Do not take our word for the limits. Three checks, five minutes, on your own idea and your own answer.

Check 1

Ask the same idea again in a fresh chat

Same prompt, same idea, new window. Compare the two verdicts. They will not match, and usually not by a little: 6/10 and 8/10 on the same paragraph is normal. Nothing in a prompt anchors the score to anything measured, so what you got was a mood, formatted as a rubric. If the second answer is better news than the first, notice how much you want to believe that one.

Check 2

Try to click through to one number

Take the market size it gave you and ask where it came from. You will usually get a plausible-looking attribution to a research firm and no link, because the figure came out of training data rather than a document. Now imagine that number inside a deck, in front of someone who checks. A confident paragraph reads exactly the same whether the source exists or not.

Check 3

Ask it to name five competitors, then check them

Search each one. Expect at least one that does not exist, and at least one real company that shut down or pivoted a year ago. Neither is a bug you can prompt away: one model in one pass has no way to catch its own invention, and no reason to hesitate before writing it.

None of this makes the prompt useless. It makes it a first pass — a way to find the questions worth answering, not the answers themselves. The failure mode is not that the output is bad. It is that the output is convincing, and you stop there.

What closing those gaps requires

Each of the three checks fails for a structural reason, and none of them is fixable by writing a better prompt.

The gap	What it takes to close it
Different verdicts each run	The verdict is computed, not written. Market size, growth, business model, problem clarity and audience each score against fixed thresholds, and the verdict is a threshold on the sum. Run it twice and the number moves only if the evidence moved.
Numbers without sources	Six research stages run with live web search, and every claim carries the link it came from. Where a figure is an estimate, the report says so.
Invented competitors	Key claims go through three independent models from three vendors plus web evidence, and when they disagree the disagreement is recorded by name rather than averaged into a smooth sentence.

And the input differs before any of that starts: instead of a paragraph you typed, a 15-minute conversation that asks follow-up questions and pushes when an answer is thin. Measured on our own projects, that brings ten times more context into the research — median 1,616 characters of your own words against 163 from a single field. The prompt above is limited by what you thought to write down. The interview is not.

Run the whole thing instead

Same task, different machine underneath: a voice interview, six research stages with live search, three models checking key claims, and a computed verdict. Three projects free, no card.

Start free — 3 projects →

15 min · No credit card · up to 20 reports

Frequently asked questions

Why does ChatGPT always say my idea is good?+

Because you asked it whether the idea is good, and it is built to be helpful. Ask "is my idea good" and you get agreement dressed as analysis. The prompt below removes the agreeable path: it forces a named verdict against fixed criteria, requires the model to argue the case against your idea first, and asks for the specific evidence that would change its mind.

Does this work in ChatGPT, Claude and Gemini?+

Yes, all three. Use the strongest model you have access to — this prompt asks for structure and self-criticism, and weaker models tend to drop half the sections. If your tool has web search, turn it on: without it the competitor section is written from memory and will be out of date.

Can I trust the numbers it produces?+

Not without checking them. A language model will produce a market size and a growth rate for any idea you give it, including one you invented five seconds ago. Treat every number as a claim to verify, not a finding. The prompt asks the model to mark which numbers it is confident about and which it is estimating — that section is the one to read first.