Thirty-six entries in, my error log stopped adding up
I keep a record of where the AI gets things wrong and what caught it. When I went back to count, half my catches might not have been catches at all.
I stopped writing rules for the AI and started building checks
My instructions file kept growing and the rules in it kept getting broken. The fix was not a better-worded rule.
I use one model to check the other one's work
Claude writes it, Codex reads it. Not because either is better, but because they are wrong about different things.
I built myself an operating model and then it told me I was wrong
Six months into a product with solid research and a working prototype, I had zero conversations with the target population. So I built a framework, applied it to my own work, and acted on what it found.
AI has preferences and I'm not sure what to do with that
I asked Claude Code to be honest about what kind of work it engages with and what it tries to avoid. The answer changed how I manage the partnership.
When one Claude Code session wasn't enough
How the evolution of a discovery project led me to stand up scoped Claude Code instances as the first step toward an agentic workflow.
Nine days to a live market test
A stakeholder said 'do a market test.' Nine days later we had persona-based landing pages, data capture, interview-informed copy, and a go/pivot/kill framework.
I built four agents to disagree with me
With no product team around you, you have to manufacture the friction a larger team provides naturally.
Speed works until it works against you
An automated file sync tool wiped an entire Drive folder. What I learned about guardrails, manual control, and designing systems where speed doesn't kill.
Three articles that changed how I work
How connecting articles to active project decisions turned passive news consumption into something I actually use.