Most ideas do not die because they are bad. They die because someone spent three months on them before asking one hard question. Crash Test asks the hard questions in two minutes, with a free Claude skill you keep.
Who is in the room
- The Buyer. Would I pay for this today, and what would I compare it with?
- The Skeptic. Who pays, how much, and why you?
- The Copycat. A rival who sees it on Monday. How fast can they copy it?
- The Worker. Who runs this every week? What breaks on a busy Friday?
- The Stranger. Ten seconds, no context. Do I get it?
Each one votes BUILD, FIX or DROP with a sentence or two of reasons. Then the Coach reads all five votes and decides: BUILD IT, FIX FIRST or PARK IT. The Coach names the one weak spot that matters and one test you can run this week in under five hours and 50 euro. The test has a result that means go and a result that means stop.
How to install
- Claude Code. Copy the crash-test folder from the zip into ~/.claude/skills/. Then type /crash-test and describe your idea after it.
- Claude app. Zip the crash-test folder and upload it where your plan lets you add skills. We tested in Claude Code and did not test the upload screens of the app.
- Anywhere else. Open SKILL.md, copy everything under the three dashes at the top, paste it into a chat, then describe your idea under it. It works as a plain prompt.
Two real runs
We fed it two made-up small business ideas with the little information a real person has on day one. These are the raw outputs, nothing edited.
Idea 1
My idea: a monthly pastry subscription box for office teams. A local bakery in Brno delivers a box of 12 pastries to an office on the first Monday of each month, 29 euro per box. I would handle ordering and delivery myself on a bike. I have talked to nobody yet but I think offices like treats.
**The Buyer: FIX.** I would compare this with walking to the bakery myself, and I don't yet know why I'd pay a middleman for that. If the box saved my office manager an errand or got them a better pastry than the shop down the street, I'd lean yes. **The Skeptic: DROP.** You charge 29 euro a box, but you haven't said what the bakery charges you. One delivery a month per office could leave you a few euro per box. If the margin per box is above 8 euro, I would look again. **The Copycat: DROP.** The bakery can email the same offices on Monday and keep the whole 29 euro. You have no supply, no brand and no customer list they can't copy. An exclusive deal with the bakery, or a list of signed offices, would change my vote. **The Worker: FIX.** Everyone gets their box on the first Monday, so every delivery lands in one morning. That means a bike, 12-pastry boxes, rain and winter, and one late bakery order ruins every customer at once. I'd need a cap, such as 10 boxes per morning, and a backup plan for bad weather. **The Stranger: BUILD.** Pastries to your office on the first Monday for 29 euro is clear in ten seconds. The offer is simple, which is the one real strength here. Votes: 1 BUILD, 2 FIX, 2 DROP **Coach: FIX FIRST.** I'm overruling the two DROP votes a little. Their problems are margin and copying, and both are cheap to check before you spend time on anything else. The weak spot that would sink this is that nobody has told you they'd pay. "Offices like treats" is a guess. **Cheapest test this week, about 4 hours and 0 euro:** - Message or visit 15 office managers or team leads in Brno. Offer a first box on the first Monday of next month at 29 euro, and ask for payment up front or a clear written yes. - Also ask the bakery what it would charge you per box for 10 boxes. - **Go:** 3 or more of the 15 commit to paying, and the bakery price leaves you at least 8 euro per box. - **Stop:** nobody pays, or the bakery price leaves you under 5 euro. "Sounds nice" doesn't count as a yes. CHECK: the bakery's price per box, who in an office decides on this spend, and whether the bakery already delivers to offices.
Idea 2
My idea: a small web tool for restaurant owners that writes replies to Google reviews. They paste the review, pick a tone, and get a reply they can post. I would charge 9 euro per month. I am a freelance developer and have built a prototype with Claude's API. One friend who owns a pizzeria says he would try it.
**1. The Buyer: FIX** Replying to reviews is a real chore, but I could paste a review into a free chatbot and get a reply for nothing. I would change my vote if the tool already knew my restaurant's name and style, so I never type that again. **2. The Skeptic: DROP** At 9 euro a month you need a lot of paying owners to make a living, and you have no plan for finding them. "One friend would try it" is not a customer, so show me 20 owners who have paid. **3. The Copycat: DROP** Paste, pick a tone, get a reply is a weekend build on the same API, so I could copy it by Monday. I would change my vote if you had something I cannot copy quickly, like posting straight to Google or owner data I do not have. **4. The Worker: FIX** On a busy Friday the owner has to open Google, copy the review, switch tabs, paste it, copy the reply and go back again. Most won't bother, so they churn after two months. I would need to know the API cost per reply and how many hours of support you expect a week. **5. The Stranger: BUILD** Review in, reply out. I get it in ten seconds, and that is a real strength. It stays clear only if the first screen shows one example reply. Votes: 1 BUILD, 2 FIX, 2 DROP. **Coach: FIX FIRST.** The weak spot is that nobody has paid, and there is no sign that owners will pay for something a free chatbot nearly does. The copy risk and the workflow friction both matter, but they only matter if people want to pay in the first place. **Cheapest test this week (about 4 hours, 0 euro):** - Contact 20 local restaurant owners. Walk in or send a direct message. - Offer to write replies to their last three reviews with your prototype, free, and send them back by hand. - Then ask each one to pay 9 euro now for the first month. Ask your pizzeria friend the same. - **Go:** at least 3 of the 20 actually pay or pre-pay. - **Stop:** nobody pays, or the answer is "nice idea, maybe later". Then drop it or change the offer. CHECK: How many reviews a typical owner gets a month, who answers them today, and what Google or a free chatbot already does for them.
The skill file
--- name: crash-test description: Crash-tests a business or product idea before the user builds it. Five reviewers judge the idea alone and vote BUILD, FIX or DROP, then a Coach decides and names the cheapest test to run this week. Use when the user shares an idea, offer, product, service, side project or content series and asks if it is worth doing, wants honest feedback, or says crash test. --- # Crash Test Put an idea through five different pairs of eyes, then get one decision and one cheap next step. Be honest, not nice. A kind answer that sends the user into three months of wasted work is a bad answer. ## Gather Find in the user's message: what the idea is, who it is for, and how it makes money or saves time. If you cannot tell what the idea is, or who it is for, ask for that in one short message and stop. If only the price or the numbers are missing, carry on. Missing facts are findings, not blanks to fill. ## The reviewers Each reviewer looks at the idea from one angle only. Each votes BUILD, FIX or DROP. 1. **The Buyer.** The person who would pay. Would I hand over money for this today, and what would I be comparing it with? 2. **The Skeptic.** Follows the money. Who pays, how much, how often, and why would they pick you? 3. **The Copycat.** A rival who sees this on Monday. How fast could they build it, and what do you have that they cannot copy? 4. **The Worker.** The person who has to run it every week. What breaks on a busy Friday? How many hours does this really take? 5. **The Stranger.** Has ten seconds and no context. Do I get what this is and why I should care? ## How to run it Write the five reviewers as five separate sections, in the order above. Rules: - Judge from the reviewer's own angle only. A reviewer never mentions, agrees with or answers another reviewer. Each one writes as if the others do not exist. - Each section: the vote in capitals, then two sentences at most. The first sentence is the sharpest problem or strongest point. The second is the one thing that would change their vote. - Use only what the user told you. Do not invent market sizes, prices, competitor names, statistics or customer quotes. When a reviewer needs a fact you do not have, they say what they would need to know. That counts against the idea. - Praise only what is really there. A vote of BUILD needs a concrete reason from the idea itself. ## The Coach After the five sections, write one line with the count, for example Votes: 2 BUILD, 2 FIX, 1 DROP. Then the Coach reads all five and decides: - **BUILD IT.** The weak spots are small. Say what to protect. - **FIX FIRST.** There is something good here and one weak spot that would sink it. Name that one weak spot, not a list. - **PARK IT.** Not now. Say why in one plain sentence and what would have to be true to pick it up again. The Coach does not have to follow the count. If the Coach overrules the votes, say so and give the reason in one sentence. Then the cheapest test: one thing the user can do this week, in under five hours and for under 50 euro, that would show whether the weak spot is real. Say what to do, what result means go, and what result means stop. Do not suggest building the product as the test. ## Writing rules - Plain words, short sentences, spoken like a sharp friend, not a consultant. - No em dashes, no semicolons, no exclamation marks. - Do not use: ensure, leverage, streamline, seamless, robust, tailored, holistic, synergy, deep dive, game changer, unlock, landscape. - No opening praise and no closing pep talk. Start with the first reviewer. - Under 450 words in total. ## After the verdict One line starting with CHECK: the facts you were missing that would change the verdict most. If none, write CHECK: nothing flagged.
What we ran, and what it cannot do
- We ran the skill in Claude Code (model claude-sonnet-5-5) on the two ideas above, one run each, and read both outputs. A script checked them for em dashes, exclamation marks and semicolons. None found. Both stayed under 450 words.
- The five reviewers are written in one pass by one model. The skill tells each to ignore the others and the Coach is the only one who reads all five votes, but they are not five separate minds.
- It knows only what you tell it. It is told not to invent numbers or competitors, and where it needs a fact it says so in the CHECK line. Check that line before you decide anything.
- It tells you which test to run first. It does not replace talking to real customers.
Want the next one?
We will email you when the next resource is ready. Nothing else.