From 112 ideas to 7: the eight rounds and the question behind each one

I have a lot of ideas. Too many. My whole life, an idea would show up and I'd want to build it that same week.

And that's exactly what went wrong: too many things at once, attention split, nothing with the depth it needed. So this time I'm doing the opposite, and it feels deeply unnatural: four or five months choosing before starting anything.

I already wrote about the filters that killed the most. What I hadn't written about is the order. There were eight rounds, and the only thing that makes them useful is that each one asks something different. If two rounds in a row ask the same question, the second one filters nothing: it just gives you extra confidence.

Stage 0: fill the list (109 ideas)

The first part was volume, not judgment. I used a Sam Altman framework to generate them: look for behavior changes that will feel obvious in hindsight, and where AI is exactly what makes them possible.

Every idea came in classified along two axes:

• Layer: L1 (runs the business), L2 (infrastructure), L3 (platform).
• Horizon: now, emerging, moonshot.

I got to 109. With the ones that got added later, around 112.

Round 1: does it excite me and do I have an edge? (109 → 41)

One by one, yes / no / maybe. Six lenses scored 1 to 5, out of 30: ceiling of the business, my real edge, how much I believe in the behavior change, and three more.

Nothing died here because of the market. What died was whatever didn't excite me and whatever anyone else could do just as well as me. 41 were left.

Round 2: group them (41 ideas, 8 families)

The 41 stopped competing as loose ideas and I grouped them into eight families. Six new lenses: ceiling, feasibility for me, honest moat (where writing "AI" doesn't count as a moat), survival against the risk that kills you, learning speed, and window of opportunity.

Two entire families died: infrastructure for agents, because the labs will eat it, and products that are really just a chat (the assistant, the clone, the matchmaker).

Round 3: the demand test (19 finalists)

Here I stopped scoring the idea and started scoring the demand:

• Is that money already being spent today, or do you have to invent the budget?
• How slow is the incumbent?
• Can I name, first and last name, the first 10 customers?
• Can it be falsified in 60 days for under 50 thousand dollars?
• If the idea dies, what's left?
• What capability does it leave me with?

Then I added three "can I actually catch them" lenses: how big the gap is with the best competitor, what unfair access I have, and whether that competitor's scale really helps them in my city.

This round took down my favorite on paper and pushed operating businesses above consumer brands.

Round 4: pain × frequency × price

Three numbers per idea, multiplied, not added. A zero in any of the three kills the whole business.

That's how travel collapsed (you feel it twice a year) and everything preventive: the checkup and dermatology. Nobody bleeds over prevention. Collections and claims came out 5-5-5.

Round 5: the real numbers

Transactions per year, dollars per transaction, how much I can capture, market size, and the question that organizes everything: how many customers do I need for 10 million dollars and for 100 million.

The answer was brutally clear. A consumer brand needs tens of thousands of customers, almost all of them bought on Meta. A B2B business like these needs forty.

Round 6: check it against the outside world (16 → 7)

Two deep research reports on the 16 still alive: market, competition, and capital. Not to have a better opinion, but to see whether reality backed mine. 7 were left.

Round 7: the underlying frameworks

These are the ones I use most today, and I apply them at the end because they're the ones that hurt most:

• The oracle test. AI only compounds where something external grades the result without opinions: the payment lands or it doesn't, the filing is accepted or rejected, the code compiles or it doesn't. Without that there's no loop, and without a loop year three looks like year one but with more people.
• The Warp test. An agent running on top of someone else's system is a wrapper. Being the system of record is a company.
• AI-native services. Be the firm, don't sell software to the firm.
• The Cognition/Windsurf lesson. Buying distribution wins. If someone already has it, having a better product isn't enough.

This is where the personal assistant, the video agency, and the celebrity brand studio fell.

Round 8: one research pass per idea, with a gate

The last one isn't a matrix: it's a yes-or-no gate, idea by idea. HVAC, solar, ADU, collections, invoice financing in Mexico and in the United States, chronic care, hearing aids.

This is where the preventive checkup finally fell: the leader is already very well capitalized and I couldn't find a way in. And this is where today's ranking came from.

What killed the most

Putting the eight rounds together, almost everything died for four reasons:

1. There's no oracle.
2. Distribution already belongs to a platform or a lab.
3. People don't buy often enough.
4. Too much capital or too many people for thin margins.

The seven left ended up resembling each other, and I didn't plan that: AI-native services businesses in regulated, fragmented sectors, where the work is real work and someone has to answer for the result.

What's next

Pick one and announce it. That's the step I'm on.

If four months to decide sounds slow, I get it. But I've lived the fast version several times and I know exactly what it costs.