Speaking the prompt instead of typing it: why VoiceNote started as a terminal script

Anyone building with AI learns early that the output depends on the context you give — and that giving context is expensive. Typing, you summarise. The summary is exactly where the assistant loses what would have made it right: the constraint you did not think worth mentioning, the attempt that already failed, the reason behind the odd choice.

VoiceNote started because of that, and not as a product. It was a terminal script: it recorded speech, sent it for transcription, gave back punctuated text. Fifteen minutes of talking to yourself became two pages — pasted straight into the session as context, or into the project doc as a draft.

What changed was not typing speed. It was the cost of being thorough. An idea arrives with no shape, and writing charges for structure up front. When describing the whole situation is nearly free, you stop pruning too early. And what reaches the assistant stops being your summary of the problem and becomes the problem.

Where this does not work. What comes back is raw thinking: repetition, "I mean", the sentence abandoned halfway. For a prompt that is fine — models handle redundancy well. For anything another person will read, there is work left. Solving not losing it is not the same as tidying it: whoever expects finished text will be disappointed, whoever expects raw material will not.

And there was a limit that only showed up later: a command line only runs in front of the computer. The idea that arrives on the street kept getting lost. That became the next stage, and it is another day's subject.

---

You are here. VoiceNote is a voice transcription service: you speak, it gives back punctuated text, ready to copy. It runs in the browser with nothing to install, and there is a free, open-source command-line version. This is the first post in a series about how it was built, in the order it happened. September is stage one: the command line.