A validation prompt is only as honest as its rules. The one below makes ChatGPT answer 12 deal-breaker questions — the same ones our own idea check asks — with PASS, FAIL or UNKNOWN, forbids it from inventing a competitor or a price, and tells it that one FAIL ends the idea. Copy it, fill in the four brackets about yourself, and read the answer the way the second half of this page describes.
The prompt
You are reviewing a business idea for one specific founder. Your job is to find the reason to drop it, not to encourage it. Founder: [hours a week you can give it], [money you can put in], [your skills], [will or won't do sales calls]. Idea: [one sentence]. Buyer: [who pays]. Problem: [what they do today instead]. Answer each question below for this founder with PASS, FAIL or UNKNOWN, plus one sentence of reason: 1. Can it make money? 2. Can it charge on repeat? 3. Can it charge enough? 4. Can you reach buyers? 5. Low hands-on support? 6. Safe from one platform's control? 7. A real product, not just an AI wrapper? 8. Can you sell and run it? 9. No heavy compliance blocker? 10. Sells without 'book a call'? 11. Pays for itself within a year? 12. Can it be sold or handed off? Rules: - For any fact about the market, prices or competitors, give a URL you actually read, or write UNKNOWN. Do not invent competitors, prices or numbers. - Do not give an overall score. - If any answer is FAIL, the verdict is "Drop it", whatever the other answers say. - If 3 or more answers are UNKNOWN, the verdict is "Not yet". - Otherwise the verdict is "Test it first". - End with the verdict, then the single cheapest test that would turn the biggest UNKNOWN into a PASS or a FAIL.
Fill in the founder line honestly. The hours, the money, the skills and whether you will get on a sales call are what most of the answers turn on. An idea that works for a funded team with a salesperson can fail several of these questions for one person working evenings, and a prompt that does not know which of the two you are will answer for the easier one.
Keep the problem line concrete. "What they do today instead" is the useful version of the question: a spreadsheet, an assistant, a tool they complain about. "They struggle with invoicing" gives the model nothing to check.
What most validation prompts leave out
Of the published prompts we could read on 2026-09-28 — the ones on the first page of results for this search that were not behind a subscription — the most demanding asked the model to be "brutally honest" and left it to decide what that means. None told it to write UNKNOWN instead of guessing, none set a condition under which the answer is drop, and none gave you a way to check the answer afterwards.
Asking for honesty changes the tone. Rules change the output. Each rule in the prompt above exists to close one specific way a chat answer goes wrong:
| The rule | What you get without it |
|---|---|
| A URL it read, or UNKNOWN | A competitor with a price, both plausible and neither checkable |
| No overall score | A number that moves when you rephrase and never says no |
| Any FAIL means "Drop it" | 11 passes outvoting the one question that ends the idea |
| 3 or more UNKNOWN means "Not yet" | A confident verdict standing on answers nobody could check |
| End on the cheapest test | An answer you agree with and do nothing about |
Why a model reaches for the warm answer in the first place — what it was trained to reward, and what the tools built on it are tuned for — is covered in why AI says your startup idea is great. This page is about the prompt's half of the problem.
How to read the answer in five minutes
Run it twice. A model samples, so two runs can disagree. A question that flips between PASS and FAIL across runs is an UNKNOWN in practice; count it as one. Our own check has the same property — its ratings move between runs too — which is why the rules, not the ratings, carry the decision.
Open one link. Pick the citation the verdict leans on hardest and open it. If the page does not exist, or does not say what the answer claims, treat every other citation in the answer as unread.
Change the founder line. Swap "won't do sales calls" for "will", or double the hours. If no answer moves, the model is not reading your situation, and the verdict is about the idea in the abstract — which is not the question you asked.
Where this prompt and our product differ
The questions are the same. The rules are simplified, and in three places on purpose:
- It has no "Build it" verdict. Our check reaches it only when several kinds of evidence are strong and the route to buyers is confirmed. A prompt has no way to hold a chat to that, so it stops at "Test it first".
- It is stricter about one question. Our gate treats dependence on a single platform as a mark-down rather than a kill. A chat has no honest way to weigh that exception, so here any FAIL ends it.
- It gives no score. Ours rates 7 areas and computes the total in code; a chat asked for a number produces a plausible one, which is the thing this prompt exists to avoid.
How it works has the full decision rules.
What the prompt still cannot do
It can only cite what it can open. With search off, the honest answer to every question about the market is UNKNOWN — and a run of UNKNOWNs is the prompt working, not failing.
It cannot see complaints nobody showed it. The problems our sweeps have recorded from public sources sit on the problems pages; a chat knows only what it remembers or finds in the session.
It does not keep a decision. A chat leaves a transcript, not a verdict that later answers are checked against, so nothing stops the same idea coming back next month and passing. WhittleOS vs ChatGPT sets the two side by side.
To run the same 12 checks with the rules enforced in code rather than requested, a Single Idea Check returns a verdict in under a minute, and the first one is free with the 2 credits every new account gets.
The short version
- The prompt asks our 12 deal-breaker questions and allows three answers: PASS, FAIL, UNKNOWN.
- Its rules do the work: cite or say UNKNOWN, no score, one FAIL means "Drop it", 3 or more UNKNOWNs mean "Not yet".
- Check the answer three ways: run it twice, open one link, change the founder line.
- It stops at "Test it first" on purpose. A chat can argue with an idea; it cannot gather the evidence a build decision needs.
Common questions
Is ChatGPT good at validating business ideas?
At arguing with one, yes: the obvious objections, the segments you missed, a first read on who might pay. At the facts under the argument, only when it can open pages — with search off, the honest answer to every market question is UNKNOWN, and that is what this prompt tells it to say. Treat it as a free second opinion, not a decision.
Can I trust the sources it cites?
Open one before you trust any. A citation is worth something only if the page exists and says what the answer claims. If the first one you open fails that, treat the rest as unread. The prompt asks for URLs it actually read, but that is an instruction to the model, not a guarantee, so the check stays yours.
How is this different from a validation tool?
The questions are the same 12 our idea check asks. What changes is who enforces the rules. In a chat the model follows them only as far as it chooses to; in WhittleOS the verdict rules and the score are computed in code, 3 or more unknowns hold an idea rather than drop it, and the decision is saved. The first idea check is free with the 2 credits every new account gets.