business-idea-finder

A search engine for opportunities, not a brainstormer.

Most idea tools generate. Generation is not the scarce resource — you can have fifty ideas in a minute, and they will nearly all be plausible, unfalsifiable and worthless. This one searches: inside what you already know, against evidence that real people left in public, through screens that kill the shapes that reliably fail, and it ranks what survives by how fast you can prove it wrong.

It is the companion to business-idea-validator. The validator answers is this good? This answers what should I do instead?

npx business-idea-finder

Then restart Claude Code and run /business-idea-finder.

Why it works this way

Ideas are not thought up. They are searched for, inside the founder's own information set.

Shane (2000) gave eight teams the same MIT 3D-printing invention. They built eight completely different businesses, and what explained the difference was what each team already knew. So the skill builds your advantage inventory first — what you know, who you can reach, what you hold, what you have already built — and searches inside it. Ideas outside it are flagged as outside it, never quietly included.

The scarce resources are discernment and cheap evidence. Girotra, Terwiesch & Ulrich showed the quality of your best idea depends on variance and on your ability to tell good from bad. So ten generators run independently, wide, without seeding each other — and then the filtering is a real stage rather than an afterthought.

Lead users leave traces. Von Hippel found lead users pioneered 54.4% of the most important innovations across 1,678 cases. People who built a workaround usually complain about having had to. The skill goes and finds those complaints — verbatim, dated, linked — before it proposes anything.

The fitness function

Not "which idea scores highest." Expected rubric points per week of testing.

An idea that could reach Promising after two weeks of conversations beats one that could reach Strong after six months, because the second will never actually be tested — and an untested idea is worth its current score.

| | Candidate A | Candidate B | | --- | --- | --- | | Predicted now | 46 | 46 | | Ceiling if the test passes | 62 | 78 | | Weeks to get there | 2 | 20 | | Velocity | 8.0 / week | 1.6 / week |

A wins, despite the lower ceiling. Every candidate therefore carries a cheapest disproof: the fastest test that would kill it, runnable in under two weeks by one person.

The six screens

Structural questions, answerable before anything is built, because the 2026 evidence shows the same businesses dying of the same knowable causes:

1. Margin structure — does cost of goods scale with revenue?
2. Billable outcome — can you name one countable unit you get paid for?
3. Price floor — under $50/month, AI-native gross retention is 23%. Over $250, it is 70%.
4. Embedded distribution — what is the structural reason people find this? "We'll do content marketing" is not an answer.
5. Verifiability — revenue concentrates where output is inspectable by the person supervising it, and vanishes where an agent acts irreversibly on someone else's systems.
6. Predator lane and enforcement — is a lab occupying this? And if the why-now is regulatory, has anyone actually been fined?

Every kill is printed with its reason. You learn as much from what died as from what lived — and the rendered page gives the kill list the same weight as the survivors, which is the opposite of what every other ideation tool does.

You get a page, not scrollback

Every run renders a shareable dashboard automatically: the funnel with its rejection rate, the survivors with their scores drawn against the band thresholds, the kill list with the screen that made each call, and — when the ranking metric and the recommendation disagree — that disagreement stated outright instead of buried.

Honest about its own limits

This skill cannot create demand evidence. It can only shorten the path to it.

Demand evidence is 30 of 100 points on the bootstrap path, and it cannot be obtained by reasoning — only by talking to people. So it is predicted at zero, always, and may never be asserted. Which means every candidate this skill produces starts at Weak, by arithmetic.

That is the design working, not a flaw in it. Holding the largest component at zero makes it un-gameable, so optimising against the rubric pushes candidates toward testability rather than toward cleverness. The 7-day plan is what buys the points.

It also carries a do-not-cite list, because most circulating "X% of AI startups fail" figures trace back to content farms, and a false-alarm list of regulations that were delayed, narrowed or repealed after vendors built for them. Regulatory dates are re-verified live on every run — the skill's own design document once asserted an EU AI Act date, taken from an aggregator, that was wrong by sixteen months.

Commands

| Command | What it does | | --- | --- | | /business-idea-finder | Full search against your advantage inventory | | /business-idea-finder <field> | Same, intersected with that field | | /business-idea-finder reframe SLUG | A different angle on an existing project's assets | | /business-idea-finder inventory | Build and print the advantage inventory only |

Install

npx business-idea-finder          # from npm
./install.sh                      # from a clone
npx business-idea-finder -n       # dry run, changes nothing

Installs to ~/.claude/skills/business-idea-finder/. An existing install is moved to ~/.claude/skill-backups/ first — never overwritten, and never left inside skills/, where a backup would register as a second skill.

Pairs with the validator, but does not require it: references/scoring-summary.md carries enough rubric to predict on its own, and the skill says what to install rather than degrading silently.

Scope

It searches and it shortlists. It does not write code, and it does not modify your projects. It writes one thing — the shortlist and prediction log, under ~/.claude/business-idea-finder/. There is no Edit and no Bash in its allowed tools, so nothing already on disk can be changed by it.

License

MIT