One Voice, Ten Lanes: A Day of Dictation

Two kids, three startups, one voice told the origin story: parallel high-stakes startup lanes, a second child under six months old, and a keyboard that had quietly become the single point of failure for everything. That post explained why Keebye exists. People kept asking a more practical question: what does a day actually look like?

So this is that day. Not a typical day, because there is no such thing, but a real shape that repeats. I am going to avoid numbers on purpose. I have never measured a percentage improvement and I would not trust one if I had. What I can describe is where the voice channel gets used, where it does not, and what changed about the texture of the work.

A note on "ten lanes." It is not a boast. A lane is anything that advances on a sentence from me: a coding agent mid-task, a code review someone is waiting on, a Slack thread, an email to an investor's lawyer, a document draft. On a busy day there really are around that many open. Most are idle at any given moment. The problem is that the ones that wake up do so on their own schedule, not mine.

Early morning: one arm, three agents

The day starts before the house does, which means it starts with a baby on one arm. This is the hour where voice went from convenience to necessity.

Three agent lanes are already running from the night before, each in its own terminal, each in a different repository. The first thing I do is read what they did. Reading needs eyes, not hands, so the baby is fine. The second thing is to redirect the ones that went sideways, and that is where a keyboard used to lose.

Now: click into the terminal, hold Right Command, say the redirect, release. "The migration is fine, but you renamed the table in the model and not in the seed script. Fix the seed and rerun the tests." Text lands in the prompt. Enter. Next lane.

I have a small habit that took a while to learn: speak in full sentences. Agents do better with them, and dictation produces them naturally because it is how people talk. The telegraphic style I used to type ("fix seed, rerun") was a keyboard artefact, not a preference.

This hour also exposes the one first-dictation quirk worth knowing. The microphone warms up after the first dictation of a session and keeps a 500 ms pre-roll from then on. The very first hold of the morning can pay a stream-start cost and clip a word if I start speaking too early. I have learned to pause a beat. Why dictation apps eat your first word explains the mechanics; the short version is that it is a category problem with a partial fix, not a solved one.

Mid-morning: the review lane

Two of the startups I work with have code reviews waiting on me most mornings. Review is the best possible dictation use case: it is reading with occasional short prose output, and the prose is addressed to a human.

I read the diff in a browser tab. When something needs a comment, I click into the box, hold, speak, release. The custom dictionary carries the project's identifiers, so reconnectBackoff comes out as reconnectBackoff and not as three English words. Cleanup is rule-based by default: fillers and repeated words go, my wording stays. I keep the optional local-LLM polish off for review comments, because the fidelity guard behind it is a heuristic against severe token loss, expansion, or repetition, not a guarantee that a negation survived, and in a review a flipped "not" changes the verdict.

One of these teams reviews in Romanian. Keebye's 25-language engine handles that on-device; I pin the language in the tray and the next utterance is decoded as Romanian, English identifiers included. Dictating code reviews in your own language has the details, including the honest list of which languages are covered and which are not.

Late morning: the human lanes

Not every lane is an agent. Around now the Slack threads and emails that aged overnight need answers, and this used to be where the day silently lost hours. Not because replies were hard, but because each reply meant leaving a lane, and leaving a lane has a re-entry cost.

Voice changed the economics of a reply. A two-sentence Slack answer is one hotkey hold, spoken into the message box while my eyes are still half on the terminal next to it. Investor-adjacent emails are longer and I still write them with care, but the first draft is now spoken, then edited. Speaking a draft gets the shape down fast; editing it typed is where the precision goes.

This is also the hour where the privacy requirement matters most. What I dictate here is the interior of other people's companies: unreleased products, security discussions, money. Keebye transcribes on-device, works with Wi-Fi off after the model download, keeps a text-only history locally that deletes itself after 30 days, and sends no telemetry. Where does your voice go is the longer version of why that was non-negotiable for me.

Afternoon: deep work, and the lanes that wake up

Afternoons are for the lane that needs actual thought: a design, a hard bug, a document that has to be right. This is typed. Voice is a control channel, not a thinking channel, and a hard problem does not get easier by being spoken.

But the other lanes do not know it is deep-work time. An agent finishes and asks a question. A reviewer replies to my comment with a counter-argument. Before, every one of these was a decision: break focus and switch windows, or let it age. Now most of them are a glance and a sentence. Glance at the agent's question, click in, say "yes, but keep the old endpoint alive until the client ships," click back. The deep-work window stays mostly intact because the interruption is a few seconds of speech rather than a full context switch into someone else's problem.

I want to be precise about the claim here. Voice does not make interruptions free. It makes the mechanical cost of responding small enough that the remaining cost is just the thought, which was always the real cost anyway.

Evening: the terminal-heavy lane

One of the startups runs a lot of its operational work over SSH, in tmux, on servers where clipboard paste does not exist as a concept. This is where dictation tools usually fall over, because most of them insert text by pasting.

Keebye pastes by default too, with terminal-aware chunking for shells. For SSH and tmux I flip the opt-in typing mode, which injects the text as Unicode keystrokes instead of relying on a clipboard. It works with non-QWERTY layouts, which matters for the Romanian keyboard on one of the machines. Dictation that survives the terminal covers the setup and the terminal guide covers the trade-offs.

Keebye also refuses to insert into secure input fields, so when a server asks for a passphrase, the hotkey does nothing. It is occasionally annoying and entirely correct.

What did not change

A day like this can read as if voice solved the problem of having too many lanes. It did not. It solved one specific bottleneck: the physical cost of delivering a sentence to whichever lane needed it. Everything else is the same.

Thinking is still the bottleneck for hard work. Reading still takes as long as it takes. Agents still go sideways and still need redirecting, and a mumbled redirect produces a mumbled result whether it is spoken or typed. Voice is a supplement to hands, not a replacement; I type plenty, just no longer as the one chokepoint that every lane waits behind.

And the app has the rough edges of early software. Identifiers get misheard. The first word of the day sometimes clips. There is no voice-command vocabulary, so navigating between windows is still a click. It runs on macOS only.

The shape of the thing

The old picture of a 10x developer was one person typing faster. The picture that actually arrived is one person conducting: several lanes, some automated and some human, each waiting on a sentence. In that world, leverage is not typing speed. It is how cheaply you can deliver intent to the lane that needs it, in whatever physical situation you are in, including one arm full.

If your day has that shape, the developers guide is the practical starting point, and if you are weighing dictation apps for it, Keebye alternatives is an honest comparison that says where each one fits better. The trial is 14 days and starts on your first desktop sign-in, so a real week of lanes is exactly what it is for.