Voice Notes Are the Killer Input

Why Voice

Typing on a phone is a tax on capture. Speaking is not. The single-tap voice message turned out to be the highest-bandwidth path into the system by a wide margin: you ramble for four minutes while walking, and the note carries context, hedges, names, and asides that a typed capture strips out. One voice note routinely becomes half a dozen tasks, a couple of vault updates, and a question or two back to you.

The Path

1. You record a voice message in the channel.
2. The coordinator’s scheduled pull picks it up like any message.
3. It downloads the audio and transcribes it locally, with an on-device speech model — sensitive content never transits a cloud transcription API. (On modern laptops this is fast enough that the cost is a few seconds per minute of audio.)
4. The transcript lands in Efforts/inbox/ as a capture, with the message ID as provenance.
5. Triage processes it like any capture — with one formatting rule that voice notes made mandatory.

Quote the Source, Annotate in Between

For any long capture, the assistant interleaves your verbatim words with the derived actions:

“…so two things on the house. The contractor said the sauna quote is coming Thursday, and I want to move the study to the smaller room — actually, ask [name] first, she had a view on that…” → Task: Contractor: sauna quote — expected Thursday; chase Friday if absent (home, waiting, followup set) → Task: Ask [name] about moving the study to the smaller room (home, @call) → Note: study-move idea recorded on the home venture hub, marked “pending her view”

Chunk by chunk, all the way through. This exists because the alternative — a tidy summary — silently dropped things. The feedback log has the entry: three named destinations, mentioned clearly in a voice note, were folded into another item’s prose and effectively lost; the reaction was, reasonably, why are you ignoring what I said? Quoting makes omission visible to the reader and therefore to the writer.

Voice as the Disposition Path

The same channel carries your replies to the brief, and voice became the primary way to dispose of items. You record one note hitting items in the order they appear in the brief: “the subscription one, cancel; the invite, decline; the lawyer thing, defer a week; the rest, fine.” The assistant parses by reference-to-section-and-item, never by task ID — you should never say an ID aloud. Every brief line is framed as a decision precisely so a one- or two-word spoken answer resolves it.

Things Voice Taught the System

• Relative day-words are ambiguous at night. “Tomorrow” recorded at 1 a.m. means the day that has already started. The triage output resolves to an absolute date and echoes it.
• Names get hedged, not guessed. Transcription mishears proper nouns. A name that does not match a people note is recorded as heard, with a “(spelling TBC)” hedge, and confirmed in the receipt.
• Numbers and spellings get read back. “N-E-E-R-A, confirmed” beats a confident wrong entry.
• The receipt lists every derived item. So you can scan it on the phone and spot the one thing that was misheard.

Outbound Voice

The reverse path exists too — a spoken reply for when you are away from a screen — with a classification gate: public content may use a cloud voice; anything personal, financial, or work-related is synthesised on-device. See Operators with narrow hands.

Getting Started

You need three things: a local transcription model (several open ones run well on a laptop), the coordinator’s pull step to download audio attachments, and the quote-then-annotate rule in your triage routine. The first voice note you send will be processed imperfectly. Correct it, and the correction becomes the format.

See it

The demo vault has two minutes of talking in a car becoming five decisions. It is fictional, and it was built by Claude Code sessions running the bootstrap prompt — not written by hand.

#voice · #capture · #transcription · #privacy