How to roast your own idea in an hour

You have an idea. You typed it into an AI. It came back with a market size, a go-to-market plan, three customer personas and the word "promising".

That took thirty seconds, and it told you nothing. The model was trained to keep you typing, and the fastest way to keep a person typing is to agree with them. So every idea gets the same warm bath.

Here is the antidote. Six questions, asked in a fixed order, each with a rule that makes vagueness impossible. You can run it with any AI you already pay for. It takes about an hour. At the end you have a verdict you can defend and the cheapest possible test to prove it either way.

We call the six questions the council. Each one is a voice with a single job and a hard rule. Give each voice its own message, in order, and do not let the next one start until the current one has met its rule.

Before you start: write the idea in three lines

Do this yourself, not with the AI. Three lines, no adjectives:

1. Who pays. One kind of person or company, specific enough that you could find ten of them on LinkedIn this afternoon.
2. What they get, in the words they would use, not yours.
3. Why now. What changed that makes this possible or urgent this year and not five years ago.

If you cannot write line 1 without the word "anyone", stop here. That is your first finding, and it cost you nothing.

Voice 1: The Wrecking Ball

Job: find the one fatal flaw and say exactly how the idea dies.

The rule: it must name one specific way the idea fails, in a sentence that starts with the failure and ends with the mechanism. "Some challenges may arise" fails the rule. "Nobody switches because the incumbent is free and already installed" passes.

The prompt:

You are the Wrecking Ball. Your only job is to find the single most likely reason this idea fails and describe exactly how it fails, step by step, as if writing the post-mortem two years from now. No balance, no upside, no hedging. One fatal flaw, named precisely. If you find yourself listing several, pick the one that kills the business fastest. Here is the idea: [your three lines]

Read the answer. If it is generic ("competition is fierce"), reply: "Too vague. Name the specific competitor, the specific price, and the specific moment the customer chooses them instead." Repeat until it is specific enough that you could argue with it.

Voice 2: The Dreamer With A Calculator

Job: find the biggest realistic upside, with the maths shown.

The rule: one quantified prize, one assumption behind it, both stated. "Huge market" fails. "If 2% of the 40,000 UK dental practices pay £80 a month, that is £768,000 a year, and the assumption is that a practice manager will change software for a 30-minute weekly saving" passes.

The prompt:

You are the Dreamer With A Calculator. Ignore every problem for one message. If this idea works, how big does it realistically get, and what is the prize most people would miss? Show the calculation: the number of buyers, the price, the take rate, and the single assumption the whole number rests on. Then name one adjacent prize that is bigger than the one the founder is aiming at. Idea: [your three lines]. The Wrecking Ball's fatal flaw was: [paste it].

This voice exists so the roast does not become pure pessimism. A verdict without an upside is just a mood.

Voice 3: The Toddler

Job: strip every assumption you borrowed without checking.

The rule: at least three assumptions found, each followed by "but why?" until it either survives or falls over. Each one must end with a verdict: survives, or falls.

The prompt:

You are the Toddler. Take this idea and list every assumption it rests on that the founder did not check themselves, including the ones borrowed from "how this industry works". For each one, ask "but why?" and keep asking until the assumption either survives with a reason, or collapses. Minimum three assumptions. End each with SURVIVES or FALLS and one line saying why. Idea: [your three lines]. Fatal flaw so far: [paste]. Upside so far: [paste].

The Toddler is deeply annoying and almost always right. The assumption that falls over is usually the one you were most sure of.

Voice 4: The Receipts

Job: no opinions. Real prices, real rivals, real numbers, with a source and a date for each.

The rule: every claim carries a link and a date, or is marked UNVERIFIED. No exceptions. If your AI cannot browse, it must mark everything UNVERIFIED and you go and check the top three yourself. This is the step most people skip and the step that decides whether your plan is evidence or decoration.

The prompt:

You are the Receipts. You have no opinions. Find: the three closest existing alternatives and what they charge today; the free thing that already does most of this; the best available number for how many buyers of this kind exist; and any evidence of people paying for something like this now. For every fact give the source URL and the date you found it. Anything you cannot source, write UNVERIFIED next to it. Do not soften, do not summarise, do not conclude. Idea: [your three lines].

Now look at what came back marked UNVERIFIED. Those are the claims your plan is currently made of.

Voice 5: The Wallet

Job: be the actual buyer, deciding whether to pay.

The rule: it must end with a decision, PAY or NO, and the single price point that would flip it.

The prompt:

You are the Wallet. You are the exact buyer described in line 1: busy, sceptical, with a budget and a boss. You have read the pitch for eleven seconds. Say out loud what you think, what you would do instead, and whether you pay. End with PAY or NO, the price at which your answer flips, and what you would need to see to trust it. Idea: [your three lines]. What the Receipts found: [paste].

Money makes people painfully honest. This voice tells you what the customer says at their desk, not what they say at your demo.

Voice 6: The Final Word

Job: read the whole brawl and rule.

The rule: one of three verdicts, GREEN LIGHT, RESHAPE or KILL, plus the cheapest test that could be run in 48 hours, with a pass mark set now, and the kill criteria: the result that would make you stop.

The prompt:

You are the Final Word. Read the five voices below. Rule GREEN LIGHT, RESHAPE or KILL, and say in two sentences why. Then design the cheapest test that could be run in 48 hours for under £100 to prove the verdict right or wrong. Set the pass mark before the test runs, as a number. Then state the kill criteria: the result that means the founder should stop. No encouragement, no summary of the idea, no hedging. [paste all five]

What you have now

A verdict, a fatal flaw, a quantified upside, a list of assumptions with the fallen ones marked, a list of claims marked verified or unverified, a buyer's decision, and a 48-hour test with a pass mark.

That is more than most pitch decks contain, and it cost you an hour and the subscription you already had.

The part the AI cannot do

Run the test. The council can design it, the market has to answer it. A price on a landing page, a deposit, ten conversations with the buyer described in line 1, a pre-order: whichever the Final Word chose.

When the result comes in, write it down next to the pass mark you set, whether it passed or failed. A failed test, logged, is worth more than a plan that was never tested. It is the one thing a beautiful deck can never fake.

Why the order matters

If you ask for the upside first, the model anchors on optimism and the flaw comes back soft. If you ask for the receipts last, the verdict has already formed without them. Fatal flaw, then upside, then assumptions, then evidence, then the buyer, then the ruling. Every time.

Three ways this goes wrong

1. You let a vague answer through. The rules exist so you do not. "Competition is fierce" is not a fatal flaw. Send it back.
2. You skip the Receipts because your AI cannot browse. Then check the top three claims yourself. Twenty minutes. The plan is worthless without them.
3. You argue with the Wrecking Ball instead of testing it. The point is not to win the argument. The point is to find out, cheaply, before the money.

---

We are building a standard for this: a public 0 to 5 rating for how well a decision has been tested, and a council that runs these six voices and remembers what you decided. If you want to be told when it opens, the list is at the bottom of the home page.