Giffen
Labs

In April I wrote about the app I couldn't ship. Proximity was too hard to test alone, iOS was hostile to the whole idea, and the music APIs didn't want it to exist. Nine days of rebuilding later it's on TestFlight with its App Store listing filled in. Almost none of the April problems got solved. They got designed out.

I ran Brown Note's numbers through Winning by Design's Bowtie, a model built for recurring revenue businesses. It made the problem obvious. Brown Note doesn't have an awareness problem. It has an onboarding problem, and nothing to do unless you're having a bad day.

Eight months after launch, Brown Note has 55 real users and about three of them show up in a given week. Two podcasts and a listing on the Crohn's & Colitis Foundation site didn't change that. Here are the real numbers, what I think is actually broken, and the redesign I'm about to do.

I pointed the newest models from OpenAI and Anthropic at the same game, a retro EHS audit RPG, and let them build, test and critique each other's work. What they were good at, where they failed, and what I'd do differently.

I finished Throughline months ago. It was technically complete, live, and quietly not landing. The fix wasn't more engineering — it was changing the question the product asks. A note on the difference between a finished feature and a product that lands.

I've been building an 8-bit app that plays a song when you're near someone you've designated a lover, a friend, or an enemy. The concept is charming. The build is stuck. Here's why proximity is the hardest product surface I've worked on, and why I'm still going to finish it.

I've been shipping three to five days of work per evening on a real project and the bottleneck isn't typing speed or context length. It's whether I spent ten minutes pressure-testing the plan before I started. The agents that stress-test my design have become the single highest-leverage part of my workflow - and the one I almost skipped the first time.

Before I bring a graph-database approach to production regulatory data, I wanted to validate the model on a domain I understand at gut level. So I built it on my music library. Here's what I learned about graphs — and why it changes how I'd architect the compliance version.

I had automated research bots emailing me competitive insights, but the information sat in my inbox half-read and never synthesized. So I built a system that never forgets - one that accumulates knowledge over months, answers questions via WhatsApp, and produces battlecards, positioning summaries, and trend analyses for other teams.

Your AI agent can read your filesystem, execute shell commands, and make network requests. You're trusting it because the system prompt says 'be helpful.' That's not security — that's hope. Here's what actual isolation looks like.