Generating ideas is cheap and getting cheaper. Eliminating them is not. A list of startup ideas is an unsorted queue with no discard policy, which is why it only ever grows.
Step one, sourcing
Stop generating and start reading. The unit of input is a documented complaint: one sentence naming who has the problem, plus a link you can open. Without the link it is something you thought while reading, not something you found.
Step two, elimination
A fixed set of questions in a fixed order, cheapest-to-establish first. The ORDER is the engineering decision — a question answerable in an afternoon of reading belongs ahead of one that needs a month of building, regardless of which is more interesting.
The measurement
From one sweep on 2026-06-04: 30 markets, 3430 candidates sourced, 93 finalists. That is 2.7%, roughly one candidate in 37. A pipeline that keeps most of what it generates is not a filter.
Two failure modes, and the first one is ours
Reading the score as a measurement. The total is computed in code from a fixed rubric, so the arithmetic is reproducible — but the ratings feeding it are sampled, so the same idea can move a few points between runs. It is an ordering. The number invites you to read it as a measurement anyway.
Gathering evidence after choosing. Evidence found before you commit narrows the field. Evidence gathered afterwards confirms the choice you already made, and you will always find it.
What we would do differently
Widen breadth before deepening one market. Raising the cap inside a single market saturates, because the stock of real documented complaints there runs out; a new market opens a fresh pool. We measured that the expensive way.
The full write-up, with the run behind every number, is at https://whittleos.com/guides/startup-ideas
Top comments (0)