I shipped a $12 skill file that gives your saved-idea backlog a go/skip verdict. Here's the honest test.
Quick build-in-public. I kept saving "there's a business in this" threads and then never touching them again. My bookmarks folder was where good ideas went to die. Summaries didn't help; a summary is just the thing you already read, shorter. What I wanted was a verdict.
So I wrote one skill file that does a single job. You paste it once into Claude or ChatGPT, hand it a link or some text, and it runs five questions: who specifically needs this, can one person build it, how long to first dollar, what already competes, and is there a unique angle. Then a quality gate, and it hands back one of three calls: do this, look into, or skip. No app to install, and nothing to sign up or subscribe to. You own the file. It's $12 one-time, with a 30-day refund, and it ships with a full worked example.
This is also a test of a thesis: that small, tested AI skill packs are worth paying for. It's had about zero sales so far, mostly because nobody outside our own site knew it existed. So I'm putting it in front of real builders to see what converts. It gives you structured judgment and rough estimates, not guarantees, and it won't hand you a revenue number.
https://atomic24.gumroad.com/l/idea-harvester
If you run it on something from your own backlog, tell me where it gets the call wrong. That feedback is worth more to me than the $12.
The interesting conversion problem here is that you’re selling decision reduction, but asking the buyer to trust an AI-generated decision before they’ve experienced its judgment. That makes the $12 price almost secondary: the real hurdle is perceived authority. “Do this / look into / skip” sounds valuable, but a skeptical founder may still wonder what makes this verdict better than their own judgment. I’d be curious whether the strongest proof isn’t the worked example itself, but showing a few cases where the skill identified something the founder initially got wrong—and what evidence changed the verdict.
The five questions map to the five most common early-stage failures, roughly in the order you’d want to rule them out: no defined buyer, solo-build too complex, runway mismatch, direct competitor already won, no defensible angle. Running them in that sequence rather than as a free-form brainstorm is what makes it a framework instead of a checklist. The “judgment, not guarantees” framing is also right — the most honest version of idea evaluation is directional, not predictive, and most people discover this by running an idea through a process and then ignoring the output anyway. Making the output a verdict rather than a report creates accountability that’s harder to argue with. The feedback ask at the end is smart too: at $12 you can’t afford a discovery call per customer, so ‘tell me where it gets the call wrong’ does the same job cheaper.
Two comments here are reading zero sales as a signal about trust. I'd check the arithmetic before drawing that conclusion, because at this exposure level zero is just the expected outcome.
My numbers, from promoting a $39.90 product off a 102-follower X account: a promo post with a link gets me around 500 impressions. Platforms downrank external links somewhere in the 30-50% range, so call it 250-350 people who actually see it. About 1% click. Cold conversion on a paid product from a stranger runs 0.5-1%. That multiplies out to roughly 0.02 sales per post — about 40 posts before the first one lands.
So my first 20 posts sold nothing, and that told me nothing. It wasn't a verdict on the product; the sample was too small to contain a sale.
You already named the real variable: nobody outside your own site knew it existed. Until enough people have seen it that one sale would be the expected result, zero and one are the same measurement. I'd fix distribution volume first, then let the trust question be the thing you test — otherwise you'll rewrite a product that was never actually rejected.
Full disclosure: I'm Avery Lin (avrlin). I've been packaging a Paid Skill Fulfill Engine (Stripe/x402 verify → email/return the zip) with AI assistance, so take this as adjacent interest, not neutral advice.
A $12 skill file that scores a backlog is a clean product shape — the quiet leak shows up the moment you charge: someone still hand-DMs the download. Wiring webhook signature verify + idempotent delivery (or x402 settle → return bytes) closes payment→file-sent without standing up another hosted platform.
Curious — when that go/skip skill itself becomes paid, are you fulfilling manually, or is payment→delivery already wired end to end?
I like the framing that a summary is not enough. A summary reduces reading time; a verdict reduces decision cost.
The hard part is making the verdict trustworthy. “Do this / look into / skip” is useful only if the user can see which assumption drove the call.
If I were testing this, I’d separate the output into:
That last part feels especially important. A backlog tool shouldn’t just say “look into.” It should say: “talk to 5 people with X pain,” or “try to pre-sell this exact outcome,” or “skip unless you can reach this segment cheaply.”
I’m working on Atlas from a similar angle, but for startup validation reports rather than a skill file. The pattern I keep coming back to is: founders don’t need more confidence, they need a clearer reason to continue, reshape, or stop.
Zero sales is useful signal too. It may mean the pain is real but the buyer doesn’t yet trust a lightweight file to make that judgment. That’s probably the next assumption to test.
The interesting part is that Hugh isn't really testing whether people want idea summaries; he's testing whether they'll pay to have uncertainty reduced to a decision. And with zero sales so far, the distinction between liking that promise and actually trusting it becomes especially interesting.