AI Guides › Workbench
By Nigel Guy · 7 min read
The usual way to use AI on a business idea is to describe it and ask what Claude thinks. You get a warm, well-structured answer, a feature list and a launch plan, and you start building. The failure is not that the answer is wrong. It is that building has become cheap enough that nothing forces you to ask whether anyone wants the thing before you spend your evenings on it.
The founder-failure statistic often quoted here ("42% of startups fail because there was no market need") comes from an older CB Insights analysis. We could not re-verify that exact figure. CB Insights' current report, published 5 March 2026 and covering 431 VC-backed startups that shut down since 2023, lists "ran out of capital" at 70% and "poor product-market fit" at 43%, with many companies citing several reasons. The direction is the same: a large share of failures involve building something the market did not pull for. Treat the number as a reason to test, not as a precise law.
The rule: never let the model judge your idea until it has been made to find evidence against it, and never build until you have a demand signal, a surviving attack and a reason the thing cannot be copied in a weekend.
| Filter | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| 1. Demand test | Turns the idea into a cheap, measurable test with a pass mark set in advance | Free Claude plan can run it; the real test costs your time | Deciding what to check this week | Claude designs the test; only real people can pass it |
| 2. Attack | Makes Claude argue the idea is a bad one, then rank the objections | Same | Finding the fatal flaw early | Claude can invent plausible objections; you must check them |
| 3. Moat | Asks what survives when a competitor can generate your product quickly | Same | Deciding what to build, if anything | Output is reasoning, not market data |
Claude's Free plan costs nothing. Pro is listed at $20 a month, about £16 at time of writing; check the £ price at checkout, and note that plan contents change. You do not need a paid plan for this stack, though longer sessions run into usage limits sooner on Free.
Claude cannot tell you whether people want your product. It can design the cheapest honest test and stop you moving the goalposts afterwards. The mechanism is a pass mark written before the test runs.
Fill in your idea, your buyer and what you can reach this week. Paste the prompt into a fresh chat.
You are a sceptical early-stage product adviser who has watched many founders confuse politeness with demand.
Context: I am thinking of building [IDEA_IN_ONE_SENTENCE] for [SPECIFIC_BUYER]. They currently solve the problem by [CURRENT_WORKAROUND]. I can reach about [NUMBER] of them through [CHANNEL]. I can spend [HOURS] hours and [BUDGET_IN_POUNDS] pounds on testing in the next [DAYS] days.
Goal: design one test that shows whether these people will take a costly action, not whether they say they like the idea.
Steps:
1. State the single riskiest assumption in my idea.
2. Propose up to three tests that fit my time, budget and reach. For each, name the action the buyer must take (pay, book a call, give a deposit, hand over data, switch from the workaround).
3. Recommend one, and set a pass mark in numbers before I run it, plus a fail mark, plus what I should do in each case.
4. Write the exact message or page copy I would send.
Format: a short table for step 2, then plain paragraphs for the rest.
Constraints: do not count compliments, likes or "sounds interesting" as passing. Do not invent market sizes or conversion rates. If a pass mark needs a benchmark you cannot verify, say so and give a range as a guess, labelled as a guess.
If any bracketed input is missing or vague, ask me for it before answering.
Before you reply, check that the test needs the buyer to give up something real and that the pass mark is a number.
Run the test, then keep the result. If it fails, that is the cheap outcome you wanted.
Models lean towards agreement. Ask for a verdict and you get encouragement, so ask for a prosecution and then a ranking.
Fill in the idea, plus the demand result from filter one (including a failed one).
You are a hostile but fair investor on an investment committee, paid to find the reason this idea fails.
Context: [IDEA_IN_ONE_SENTENCE], aimed at [SPECIFIC_BUYER], priced at [PRICE]. Demand test result so far: [RESULT_OR_NONE_YET].
Goal: produce the strongest honest case against building this, so I can decide with open eyes.
Steps:
1. List the five most likely reasons it fails, ordered from most to least damaging.
2. For each, say what evidence would prove it wrong and how I could get that evidence in under a week.
3. Name the one objection I am most likely to wave away, and why.
4. Say which single objection, if true, should stop me, and what I would see if it were.
Format: a numbered list, each item in three short lines: objection, evidence that would disprove it, cost of checking.
Constraints: no generic startup advice; every point must refer to my idea. Do not soften the case. Do not claim facts about competitors or markets you cannot support; mark anything uncertain as "check this". If you need details about the buyer or the price to be specific, ask first.
Before you answer, check that each objection is specific to this idea and that nothing is padded with praise.
Then do the checking yourself. Anything marked "check this" is a lead, not a finding.
If a competent person with Claude can rebuild your product in a weekend, the code is not an advantage. Candidates for something durable include access to a specific audience, proprietary data you legitimately hold, a workflow people already live in, trust built over time, and distribution you already own.
You are a strategy adviser who assumes any software feature can be reproduced quickly with AI assistance.
Context: my idea is [IDEA_IN_ONE_SENTENCE] for [SPECIFIC_BUYER]. My real assets are: [AUDIENCE_OR_CUSTOMERS_I_ALREADY_HAVE], [DATA_OR_EXPERTISE_I_HOLD], [DISTRIBUTION_I_OWN]. The nearest existing alternatives are [ALTERNATIVES_I_KNOW_OF].
Goal: find what, if anything, would still be mine after a rival copies the product.
Steps:
1. Describe how a rival could copy the product and how long it would plausibly take.
2. Go through my listed assets one by one and say whether each is a real advantage or just a feature, and why.
3. Name what I should build first to deepen the advantage that is real.
4. If nothing is defensible, say so plainly and suggest a smaller, honest version (a service, a niche newsletter, a paid pilot).
Format: a short table (asset, real or not, reason), then three short paragraphs.
Constraints: be blunt. Do not invent competitors or their features. Do not treat "better design" or "AI-powered" as an advantage. Ask me for missing assets rather than assuming them.
Before you answer, check that every advantage you name is something I could not get by simply paying for a tool.
Imagine a founder wants a Slack bot that summarises meetings for small UK recruitment agencies. Filter one picks a paid pilot: five agencies, £50 deposit each, pass mark of three. Two pay; the test passes, narrowly. Filter two ranks "agencies already use the note-taker built into their video-call software" first. Checking shows three of the five do. Filter three finds the code is easy to copy, but the founder's ex-colleagues in recruitment are a real asset, so the plan becomes a niche product for that audience, built after more deposits arrive. The stack did not kill the idea. It changed what got built first.