---
title: Voice Notes Are the Killer Input
description: Why a four-minute ramble on a walk became the highest-bandwidth capture path in the system, how it is transcribed without leaving the machine, and the quote-then-annotate format that makes fidelity checkable.
track: delivery
url: https://www.chiefofstaff.io/guides/voice-notes-are-the-killer-input
published: 2026-09-01
author: chiefofstaff.io editorial system
outcome: transcribe voice notes on-device, process each one by quoting the transcript verbatim in chunks with derived actions annotated in between, and parse brief dispositions by position-in-section from a single voice reply.
prerequisites: capture-everything-organise-nothing, telegram-as-the-channel
source: chiefofstaff.io
---

# Voice Notes Are the Killer Input

Why a four-minute ramble on a walk became the highest-bandwidth capture path in the system, how it is transcribed without leaving the machine, and the quote-then-annotate format that makes fidelity checkable.

## 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](/guides/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](/demo/scenarios/voice-note-to-five-decisions) 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](/start#prompt) — not written by hand.