Journal
Notes from the studio. What I shipped, what I learned, what's next.
DeskRhythm heads to beta, Fazon gets a coaching stack, and the web repo gets a quiet sweep
Week 24: DeskRhythm beta-ready with Fastlane in progress, Fazon ships time-aware coaching and Power Walk sessions, and the studio site gets maintenance.
Fable 5 replays the matrix: slower, hungrier, and the cleanest cycle yet
Run 9: Claude Fable 5 on the same iOS feature. Slower and ~3× the quota — but it internalized half my workflow and delivered the matrix's first clean fix cycle. Switching? Not yet.
Many small agents or one big one? Measuring AI orchestration on a real iOS feature
Nine replays of one iOS feature — many small agents ran ~5× slower and ~7× pricier than one big agent, with no quality edge the review gate didn't already catch.
DeskRhythm goes native, and a full week on Fazon
DeskRhythm's local-first web MVP grew a native iOS companion this week — now a TestFlight build — while Fazon added food-log export, a profile/settings home, and Trends polish.
The food tab grew up, and Product Hunt sent friends
Most of the week disappeared into Fazon — five PRDs and a follow-up landed the food + meals stack. Separately, I launched My Next on Product Hunt and discovered who actually reads new submissions.
Five quiet weeks, Fazon, and the cross-app subscription that finally works
Notes on the stretch since the My Next 1.0 launch — a deliberate step back, a new project kicking off, and the cross-app subscription that took an entitlement service to actually ship.
Agentic engineering with Agent Dashboard
Running a studio with an AI agent fleet goes faster than I expected — but the real win isn't speed. Notes on spec-alignment, token costs, and parallel-execution experiments.
Adding Claude Design as an option
Agent Dashboard had two design paths — free design via the frontend-design skill, or Pencil mockups. Claude Design became the third, then the default.
Shipping the studio site, week one
A week's worth of PRDs on a single repo. CSP, a hero rewrite, trailing slashes, route groups, the Journal itself. The hardest part was what I cut.
Mixing Claude and Codex per agent
Hitting Claude usage limits forced a per-pass model choice — implementation and review can run on different models. Multi-agent software development, one decision per pass.
Swapping Plausible for OpenPanel across the suite
Plausible was the studio's pageview layer everywhere. I swapped to OpenPanel in a day — cost was the trigger; instrumentation headroom was the why.
Where Agent Dashboard came from
The loop that became Agent Dashboard: vision plus backlog, small PRDs, multi-agent review and execution. PRD-driven development at studio pace, agentic engineering in practice.