03 / BUILT WITH AGENTS · ~2 min
I design by building with agent harnesses. These are the artifacts, live where possible, with honest notes on what I made and what I directed.
20+ Claude agents that plan my actual life: 10 city models, finance and career trackers, and a weekly synthesis run on my real decisions.
I designed the architecture, the scoring model, and every agent's prompt. Claude wrote the pipeline scripts and dashboard code; I reviewed the diffs.
[NEED: Life OS demo URL] for the redacted live demo button.
Pick fonts, colors, and component styling; it generates an agent-readable design system prompt with tokens and Tailwind components.
I designed the flow and the token schema, then directed the agent build and reworked what fell short.
This site's capybara assistant. It answers questions about my work from a curated, public-safe knowledge base.
I decide what it knows and where it stops (no personal data, ever). Claude runs the chat.
Astro, hand-tuned design tokens, a capybara scroll companion, reduced-motion fallbacks throughout.
I designed the system and the motion rules. Claude wrote most of the code; I reviewed and corrected it. Details in the notes below.
A vibe-coded travel demo: give it three places you love and it curates where to go next.
One evening of prompting. I designed the concept and the flow; the agent generated the code.
Stickers, tees, and an art label at mushimoo.com; 427+ orders on one t-shirt redesign alone.
No agents in these. I drew every sticker myself. The control group.
The agents conflated "has a deadline" with "urgent" and escalated shopping sales next to visa deadlines. The fix was a written policy every agent obeys: urgent means irreversible and tied to a value; everything else gets a badge, not a push.
Agents ended runs with "see this file path" while I read everything on my phone. A path you can't open is not a deliverable, so the rule is now: render the content, never just the pointer.
Even a bad idea gets a confident scoring model behind it, and the first version flattered me. Rankings are now blocked until a separate agent audits my reasoning first.
This page was written by me and built with Claude Code. I describe intent, it touches the files, and I review every diff.
When the output is slop, I say so and we go again. I also sell this sticker.
