I pushed back. I lost the argument. I built a few.
Here's what I learned from those hands-on experiments.
The Viral Demos Don't Survive Contact with a Client #
The first thing that caught me off guard was how fragile those viral demos really are. I tested every circulating prompt-and-skill setup I could find — the ones that dominate threads on every AI platform. One-shot output looks spectacular in a tweet or a Hacker News comment. Nobody in those threads ever has to revise the thing. The moment a client asks to change a single animation timing, swap out a section, or adjust the scroll trigger, the output falls apart. You're re-prompting, re-generating, re-editing blobs of code that were never designed to be touched. The impressive demos are impressive precisely because they were never asked to do real work. That's the gap between a proof of concept and a deployment-ready asset.
The Data Nudged Me #
The second surprise hit closer to home. Dwell time and conversion rates on the scrollytelling pages I shipped actually outperformed the minimal static pages I would have argued for. I know the sample size is small — I've been burned
Next HyperProbe: Debugging Prod Without the Redeploy Dance →