The original note, as written at the time
Texting should be for input and voice commands should be for editing and fine-tuning. That way you can synchronously create while perfecting your work, with AI.
Ongoing development
Texting should be for input and voice commands should be for editing and fine-tuning. That way you can create and perfect at the same time, with the machine.
The division is precise and it is the right way round, which is why almost nothing does it. Current systems put voice on the input side and text on the correction side, and both are the wrong assignment.
Typing is a serial, deliberate act. It is how you say the thing you already know. Voice is fast, interruptible and good at instructions about a thing that already exists: make that shorter, move that up, say this differently, read it back. Speaking an instruction is three seconds; typing the same instruction is fifteen.
Put them together and you get a working rhythm. Type a paragraph, then speak the corrections while still looking at it. No mode switching, no reaching for a menu, no reformatting the sentence by hand.
The reason it fails in practice is that voice input takes over the whole interface. You say something, it transcribes, and now you have a transcription problem as well as an editing problem. The fix is to restrict the voice channel to a vocabulary of operations rather than free dictation — a command grammar, small and learnable, with the option to see what was heard before it is applied.
The other half is that the machine should be able to ask. If an edit is ambiguous — shorter by how much, move it where — it should ask in one syllable rather than guess.
Writing is not one act. It is a draft and then twenty small decisions, and they want different instruments.