I’m rebuilding my AI startup product around one question: “What should I do next?”
I originally built an AI startup audit.
The idea was straightforward:
You give it information about your startup → AI analyzes the product, validation, risks and launch readiness → you get a score and recommendations.
The MVP worked.
But after using it myself, I started questioning the premise.
A founder doesn't really need to know:
“What's my startup score today?”
They need to know:
“What should I do next?”
And that is a very different problem.
A startup is constantly changing.
You launch something.
Users give you feedback.
An experiment fails.
Your conversion rate changes.
You change pricing.
A competitor enters the market.
You discover that one of your assumptions was wrong.
A static audit doesn't really account for that.
So I'm rebuilding my product, Plavtora, around a continuous decision loop instead of a one-time audit.
The basic idea is:
Reality changes → evidence accumulates → your beliefs change → your priorities change → you make a decision → new evidence comes in.
The interesting technical problem isn't actually generating an AI recommendation.
It's maintaining enough context about the startup that the recommendation changes when reality changes.
I'm still figuring out the right implementation.
For founders here:
Have you ever built something that worked technically, but later realized you were solving the wrong version of the problem?
This is an interesting shift. I think the difference between a useful recommendation and a generic AI suggestion really comes down to how well the system understands what changed since the last decision.
The “evidence → beliefs → priorities → decision” loop is especially interesting. How are you thinking about preserving that history so the AI knows why a priority changed, not just what the current priority is?
Exactly. I’m treating the decision history as a first-class part of the product rather than just storing the latest state.
The idea is to preserve: what changed → what belief that affected → what priority changed → what decision was made → what happened afterward.
That way Plavtora can eventually say not just “do X,” but “do X because this evidence changed this assumption.” That context is what I think makes the loop actually useful.
Yeah, treating decision history as a first-class part of the product makes a lot of sense.
I especially like the “what happened afterward” part. Without that, the system can preserve why a decision was made, but it can't really learn whether that reasoning held up.
I’m curious how you’re thinking about the messy cases though — when the same evidence later becomes stale or a priority changes for a completely different reason. That seems like it could get interesting once the history starts getting large.
The shift from score to next action feels right. A score can make a founder feel informed without making the company move.
I reckon the hard part is making the recommendation show its working. If Plavtora says “do this next,” I’d want to see which evidence changed, which assumption weakened, what option it rejected, and what risk I’m accepting by following it.
Otherwise the product can become a confident advisor when what founders really need is a better decision record. The valuable output may not be just “next action,” but “next action, because this belief changed.”
This is exactly the failure mode I’m trying to avoid. A generic “AI coach” can always generate another plausible-sounding action, but that doesn’t mean it has a reason to recommend it.
I’m leaning toward making the evidence → changed belief → next action chain explicit in the UI, including what was rejected and what risk is being accepted.
Starting with one narrow loop also makes sense. Activation/pricing are probably better candidates than trying to model the entire company from day one.
That chain would make the product much more grounded.
I’d be tempted to make the first loop painfully narrow and measurable. For example: activation. Evidence comes in, one belief changes, Plavtora recommends one action, and then the founder records what happened after trying it. Without that last step, the system can recommend forever without learning whether its advice was useful.
The rejected-options part is important too. Founders rarely choose between “do something” and “do nothing.” They choose between several plausible actions with limited time.
If Plavtora can become the place where those tradeoffs are recorded, not just where AI gives advice, that feels like a stronger product.
This feels closer to how founders actually work. The tricky bit I'd watch is whether the next action is tied to the evidence that changed, otherwise it can turn into a generic coach. I'd probably start with one narrow loop, maybe pricing or activation, and see if founders follow the recommendation before expanding the context.
Agreed. That’s probably the biggest thing I’m trying to get right with V2.
I’m thinking the first version should force a very explicit chain: new evidence → changed belief → priority → next action rather than letting the model freely generate advice.
Activation is looking like the first narrow loop to test. If founders actually follow the recommendations and the resulting evidence improves the next decision, then there’s something worth expanding.
The move from a startup score to “what should I do next?” is the real product change, not just a new prompt.
I’d make the first slice one decision loop: one evidence entry, one changed belief, and one next action, with a checkpoint before the AI writes the recommendation. Keep the builder sketch and local Git versioned the same day so the context model can change without losing the baseline.
Free Pyramid Reality Check: https://durablefoundations.gumroad.com/l/pyramid-reality-check
What single decision would Plavtora own first?
Kael Voss / DurableFoundations
That’s a good framing. I think the first slice should be one complete decision loop rather than trying to model the whole startup at once.
I’m leaning toward activation: one evidence entry → one changed belief → one priority → one next action, with the outcome feeding back into the next cycle.
The main thing I want to test is whether founders actually change what they do because of the recommendation, rather than just agreeing that the AI’s analysis sounds reasonable.