AI-Assisted Development
The constraint on building software was never typing speed. It was knowing what to build, breaking it into work that can actually be executed, and being able to tell whether what came back is correct. Coding agents removed one of those constraints and left the other two exactly where they were. Everything below is running in production.
How I use it
FocusHouse — a streaming product end to end
The brand, the product and the software beneath it. Continuous audio streaming, channel state, and a catalogue generated rather than licensed — built and shipped as a working product at focushouse.fm.
Made To Parlay — a cross-platform native app
A React Native application on Expo against a Hono API. Native mobile brings problems a web build does not have — store requirements, device state, offline behaviour — and none of them are solved by generating more code.
Specification is the actual bottleneck
What an agent produces is bounded by the precision of what it was asked for. Most of the work is upstream: deciding the data model, the boundaries between pieces, and what "done" means in terms specific enough to check against.
Reviewing code you did not write
The skill that matters is reading a diff and knowing what is wrong with it. Agents produce plausible code confidently, and plausible-but-wrong is the failure mode — it passes a glance and fails in production.
Verification in the pipeline, not by eye
Typecheck, lint, build and content-compliance checks run before anything deploys on my projects. When the volume of generated code goes up, manual review stops scaling and automated gates are the only thing that still catches regressions.
What I use
- Claude Code
- TypeScript
- Next.js
- Expo
- GitHub Actions
Where you can see it
Related skills
Want the detail behind any of this?
I can walk through any of it in as much depth as you want.