Show Notes
Day 27 of 100. There's no win to report in this one.
I'm building an AI product in public, and right now I'd rate it a 4 out of 10. That's my own score, and I'm hard on it, because I advise companies on this stuff for a living.
The product is simple to describe. Give it your website URL, and it should hand back recommendations that blow your socks off. Sales-led company? It should spot what you could be giving away for free. Already product-led? It should look at your pricing page and your signup flow and find the real opportunities. All of it in under 60 seconds, because that's one of my success criteria.
Getting there has been a tour of every AI building tool there is. Lovable got me the first 80% fast and then fought me for every point after that. I ported it to GitHub and opened a codebase I couldn't reason about, so I started over. Claude Code produced good recommendations but took 10 minutes to run. Cursor handled the file structure better, so that's where the real app got built, and we shipped it live on Render. Faster. Still not good enough. And now I get to watch what every free analysis costs us in model calls.
None of that is the actual problem.
The actual problem is codifying expertise. If you've spent years building pattern recognition in your head, getting it out of your head and into a product is the billion dollar question. I thought it was a rubric problem. I started with 30 questions and eventually cut my way down to the 13 that really matter. One of them: on your pricing page, can someone understand what they'll be charged in five seconds or less? That one is easy to write down.
Then the nuance eats you alive. A sales-led company doesn't have a pricing page at all, so what's the recommendation now? Multiply that by every edge case and you start to see the shape of it.
I've built a second app whose only job is to be the brain. I feed it tech websites and it finds the opportunities. It's the closest I've gotten, and I still haven't cracked how to break that thinking down and train the AI on it properly.
Which is why this episode is also an ask.
IN THIS EPISODE
(00:48) What the analyzer does, and why I score it a 4 out of 10
(02:09) Lovable gets you 80% there, then it fights you
(03:09) Cursor, Render, and the cost of every free analysis
(04:16) The real wall: how do you codify what you know?
(04:40) From 30 questions to 13, and the five second pricing test
(05:24) The edge cases that break the rubric
(05:45) The second app I built to be the brain
(06:17) Calling in a favor
(07:09) What we actually plan to monetize
MENTIONED
Lovable, Claude Code, Cursor and Render, the tools behind the three rebuilds
Wes Bush on LinkedIn: https://www.linkedin.com/in/wesbush/
Have you tried turning your own expertise into a product? Tell me where you got stuck.
More from ProductLed Podcast
Day 55, and I'm past the halfway point of building an AI product in 100 days. This one looks back at the first 50 days and lays out the plan for the last 45.
In the last episode I told you I'd hit a wall building an AI product in 100 days. This is what happened after I asked for help.
The podcast has been quiet for a couple of months. This episode is why.
After spending years building unvalidated products that went nowhere, Tibo Louis-Lucas completely changed how he approached startups. In this episode of the ProductLed Podcast, he shares how those early failures pushed him toward a faster, revenue-first way of building, one that eventually led to the success of Tweet Hunter and Taplio, and now powers a growing portfolio of product-led SaaS businesses.
In this episode of the ProductLed Podcast, Wes Bush and Esben Friis-Jensen sit down with Roeland Delrue, CEO and co-founder of Aikido Security, to unpack how the company reached $40M+ ARR in just three and a half years in one of the most sales-heavy categories in software.