PMF Signal

I kept taking ideas to AI tools and getting the same answer back: this has real potential, here's how to make it stronger. Every time, for every idea, including the bad ones.

That isn't a prompting problem. You can't fix it by asking for brutal honesty — the model will agree to be brutal and then find a way to like your idea anyway. It's structural. These models are tuned on human feedback that rewards agreeableness, and founders are the worst-exposed group: selected for conviction, and asking about my idea in possessive language the model reads as a cue to defer. As Inc. put it earlier this year, "AI is often very good at telling you what you want to hear."

So I built the thing I wanted to exist. Sixteen sections of live research on one idea, with three rules that aren't prompt instructions:

• The verdict is scored, not written. Pain, timing, market ceiling and founder fit are scored independently before any of the prose exists, so the number constrains the writing instead of the writing picking the number.

• Unsourced claims get cut. Every report opens with what was actually reachable. Thin sourcing is reported as a finding, not smoothed over.

• It's judged against my real constraints — actual budget, actual hours per week, actual skills — not a generic founder with unlimited runway.

The verdict is allowed to be BUILD, PIVOT, or PASS.

Then I did the obvious test and pointed it at itself.

It came back PIVOT, 5.7. Whitespace scored 3 out of 10. Under "revenue ceiling" it wrote: $50K–$300K ARR (as a product) · $0 (as the personal tool it is). Then it listed why, and none of it was comfortable:

• The incumbent is free-at-the-margin and already installed — $20/mo deep research ships with five major assistants, so I'd be asking for a second subscription to do what people already have.

• A better-funded competitor already died here. GummySearch had real revenue at $29–$199/mo and wound down commercially in November 2025 over exactly these API economics.

• The moat is a prompt. "The differentiator is a report template and a profile file."

• The idea demands distribution 9/10. I score 5.

That last one is why I'm posting this here, which I recognize is a slightly absurd sentence.

I shipped it anyway, and I changed one thing first: the subscription is gone. The report's hardest objection was that $20/mo deep research already ships with five major assistants, and asking someone for a second subscription to do what they already have is the worst sale in software. It was right — I'd had zero subscribers, ever. It's one-time credits now, $9 a report.

The rest I kept, for a reason the report also named: it's built, it costs almost nothing to run, and I score 5 on the one axis it says this idea demands a 9. The cheapest way to fix a distribution problem is to actually distribute something. This is the something I already have.

Of the reports published so far, 5 of 17 came back BUILD. The rest were PIVOT or PASS, unedited, including the one about itself.

The open question I still don't have a good answer to: does a tool whose whole value is telling you no have a market, or is "I want the truth about my idea" one of those things everyone says and nobody buys? The report thinks it's the second one. I'd genuinely like to be wrong, but I'd rather hear why I'm not.