What's this about?
dotey walks through his workflow: he prototypes with a local design-system skill, and when adding or changing a feature he changes the prototype first. His worked example is adding agent functionality to BaoCut v2 — a decision he admits contradicts his earlier position — where the agent runs on the user's already-installed codex or Claude Code CLI rather than a harness he builds himself. His prompt is unremarkable and that is the point: rethink this prototype from the user's perspective, especially how a new user gets started; don't be limited by the current design; promote Agent to the left-hand main entry with session history and recent projects; wire it into the existing translation and subtitle features; remove the agent tab from the right. He reports the first round already produced something good.
Why change the prototype before the code?
Because the prototype is the cheap place to be wrong. An HTML mockup can be regenerated in a minute and thrown away without cost, while the same exploration in the real codebase leaves debris. And models are markedly better at generating HTML layouts than at restructuring an application — so you get the model's design capability where it is strongest, then implement a design you have already looked at.
Why is it worth noticing?
- "Change the prototype first" is a discipline that predates AI and is much more affordable now.
- He explicitly notes he is contradicting his earlier position on adding an agent, which is worth more than the workflow.
- Not building his own harness — using whatever CLI the user already has — is a defensible architectural choice more people should consider.