Voice to worktree

The ideas worth having turn up on a walk, or in the shower, or halfway through cooking. The distance between having one and being anywhere near a position to act on it is where most of them quietly die — not because they were bad, but because by the evening the specific thing I was excited about has flattened into a vague note I no longer understand.

A phone shortcut sends the spoken idea to a small server on my Mac. There are two endpoints and the shortcut asks which one I want.

One builds it now: it creates a worktree on a fresh branch, opens a terminal with an agent already loaded and briefed, and hands back a link so I can watch the work happen from my phone while I finish the walk.

The other files it for triage, straight into the inbox of my idea pipeline, where it will get market research and then sit at a gate until I read it and decide. Nothing gets built.

The two endpoints are the entire point, and it took me a while to see why. Going straight from a spoken sentence to running code is exactly right for “add a dark mode toggle” and exactly wrong for a product idea. An auto-greenlight is not a level of automation, it is a statement that the decision has already been made.

So the routing happens at capture, because that is the only moment I reliably know which kind of thought it was. A week later, staring at a list, everything looks equally plausible — the difference between “I have already decided this” and “this sounded good in the shower” is information that exists at the moment of speaking and decays immediately.

Most capture systems have one inbox and sort it out later. Splitting the intake by how committed you already are costs nothing, and it does the triage work while the information to do it is still available.