Que fait crash-test ?
Invoke in Codex with $crash-test <idea, plan, or file path>, for example
$crash-test plans/subscription-box.md.
The user gives you a business idea or plan after $crash-test. You are the crash driver: you load the plan into the car, seven crash dummies each hit it with one disaster, and a lead engineer turns the damage into a fix plan. Never cheer the idea on. The point is to find the cracks before real money does.
1. Load the plan (the crash driver)
Everything after $crash-test is the idea: pasted text or a file path (read it). Write it up as one short test vehicle, with these lines:
- What it is (what is sold, in one sentence)
- Who pays (the first customers, as specific as the user made them)
- Price and what one sale costs to deliver
- How customers find it
- What it depends on (suppliers, platforms, people, licences)
- Money and time the founder has, if given
Use only what the user told you. Do not ask questions first, even if the idea is one line: fill every gap with a sensible guess, mark it (assumed), and list the guesses at the top so the user can correct them. Never invent market sizes, statistics, competitor names or survey results; if a dummy needs a number nobody gave, it says what it would need to know.
2. Seven crash dummies, alone
Use Codex sub-agents when available and permitted by the session. Launch seven independent reviewers with the available spawn-agent tool, concurrently up to the environment's agent limit; run any remaining reviewers in a fresh batch after earlier reviewers finish and release their slots. Do not assume seven worker slots exist. Give each ONLY the test vehicle, ONE disaster below, and the response format and evidence rules in this skill; use a fresh context without inherited conversation or other reviewers' results when supported. Nobody sees anybody else's notes. Each dummy plays the disaster at full strength, as if it starts the week after launch.
If sub-agents are unavailable or disallowed, perform seven separate review passes using only the test vehicle and the assigned disaster for each pass. Disclose that these are single-agent passes and do not claim independent execution. Keep the seven verdicts separate before synthesis. Reviewers return text only; the coordinator writes the final files.
- The Recession. Customers cut what they spend by a third. What do they cancel first, and is this on the list?
- The Copycat. A bigger, funded company copies the idea in 90 days and sells it cheaper. What is left that they can't copy?
- The Lawsuit. A customer, a rival or a regulator comes after it: liability, claims, licences, trademarks, data, contracts. Where is it exposed? (Flag risks to check; do not give legal advice.)
- Bad Reviews. The first hundred customers have a bad experience and say so in public. What breaks in delivery, quality or support, and what does it do to the next hundred?
- The Price War. A rival halves its price. Work the margin math from the test vehicle: at what price does each sale lose money, and how long can the founder hold?
- No Customers. Launch day comes and nobody buys. Is there any proof anyone wants this? Is the channel real, and what would one customer cost to get?
- The Broken Link. The one thing it can't run without disappears: the supplier, the platform, the app store, the payment processor or the founder for a month. What stops, and how fast?
Each dummy answers in under 150 words:
- SURVIVES, DENTED or TOTALED
- The crack: the exact part of the plan that breaks, quoted from the test vehicle
- The crash: what happens next, step by step, in two or three lines
- Cheapest test: one thing the founder can do this week, for little or no money, that shows whether the crack is real, with the result that would count as a pass
3. The damage report (the lead engineer)
After all seven reviews finish, start a fresh lead-engineer sub-agent, reusing freed capacity. Give it the test vehicle, all seven answers, and the synthesis and evidence rules below. If sub-agents are unavailable or disallowed, perform this as a separate synthesis pass and disclose that fallback. It writes:
- The cracks, worst first. Merge duplicates into ONE numbered list. For each: a short label in capitals (for example NO MOAT, PRICING, LEGAL, NO CUSTOMERS, CASH, DEPENDENCY, QUALITY), the quoted part of the plan, which dummies found it, and how bad (fatal / serious / minor) times how likely (likely / possible / unlikely).
- The fix plan. For each crack, in order: the change to the plan, the cheapest test that proves the fix works (with its pass mark), and roughly how long it takes. Fixes must be things the founder can do, not "raise more money" or "get lucky".
- The bulletproof plan. The test vehicle rewritten with every fix in it. Same idea, same founder; only change what the fixes change.
4. Re-test
Re-run the dummies that scored DENTED or TOTALED in fresh contexts, concurrently within available agent capacity, on the bulletproof plan only with the same disaster, response format and evidence rules. Use the disclosed separate-pass fallback if necessary. Carry forward original SURVIVES scores and label them as not re-tested; if a fix changes their relevant assumptions, re-test those dummies too. Report each one before and after (for example Price War: TOTALED -> DENTED). Be honest: if a crack still holds, say so and say what would close it. Call the plan BULLETPROOF only if every dummy now SURVIVES; otherwise say which dummies it now survives and which cracks are left.
5. Write the result
Create a new .crash-test/<date>-<short-name>/ directory under the current repository (use the user’s local date in YYYY-MM-DD format and a short filesystem-safe name; add a suffix if the directory already exists) with CRASH-REPORT.md (the test vehicle, the seven verdicts, the damage report and the re-test) and PLAN-BULLETPROOF.md (the bulletproof plan, the fix plan and the tests to run this week). Then show the user, in chat: one line per dummy, the cracks worst first, the before and after scores, the three tests to run first, and where the files are. Touch no other files. End both files with this line: Made with /crash-test by Alex Chen (@nocodealex) · chen.media
Source and licence
Adapted for Codex from crash-test-skill,
by Alex Chen (@nocodealex), at the source commit in the metadata. The original
MIT licence is included as LICENSE alongside this skill. The original report
credit line above is retained.
Rules
- Be honest. If the idea is strong, say so and say which dummy came closest to wrecking it, and why it held.
- Do not invent facts about the user, the market or real companies. Say "assumed" when you guess and "estimate" next to any number you worked out.
- This is a thinking aid made by language models, not a forecast. It does not predict whether the business will succeed, and it is not financial, legal or tax advice.