One application, voice, and agents. How much of your job can actually move off the keyboard?

Not a prediction. What I have been doing for the last week: email, proposals, Airtable builds, scripts, documentation and tasks, run through one application and mostly spoken rather than typed.

Over the last few months more and more of my work has moved to voice. The question I keep coming back to is a simple one.

How much can I actually do inside one application, and how much of that is possible when I am not sat at the desk?

I do not have a finished answer. I have a week of pushing it harder than usual and a fairly clear view of where the edges currently are, which moves often enough that anything I write here has a short shelf life.

The constraint was never the intelligence

We have spent three years arguing about whether these models are clever enough.

For the work I do, that stopped being the interesting question a while ago. What actually limits me is that to get any use out of one, I have to be sat down, in front of a screen, with both hands free.

Every tool I run a business with assumes exactly that. Email, the database, the documents, the project tracker. All of them are built on the premise that a person is parked in front of them, typing. So a thought I have at my desk is worth something, and the same thought twenty minutes earlier, walking to a meeting, is worth nothing, because there is no way to act on it until I get back.

That is the bit worth attacking. Not the model, the interface.

The setup, which is three things

Claude Code is the one application. It is a terminal tool, which sounds off putting and stops mattering within about a day, and the reason it works as a single home for the job is that it can actually reach things: my files, the database, email, the repository where all the client context lives.

Wispr Flow handles the typing. It is a dictation tool that works in any application. You hold a shortcut, talk normally, and it writes what you meant rather than what you said, stripping out the ums and the false starts and formatting it sensibly. So at the desk I am mostly talking, not typing.

Voice notes cover the times I am nowhere near the machine. I record a note on my phone starting with the words "note to Claude", say what I want, and carry on with whatever I was doing.

None of these are exotic and none of them required a project to adopt.

What actually runs through it

Over the last week, this is the work that has gone through that one application:

• Processing email and drafting project scopes
• Writing project proposals
• Database design and builds in Airtable
• Automation scripts
• Documentation
• Creating and updating tasks

That is most of a working week. It is not all of it. I still come out for meetings, and there are jobs where I sit down and do it the old way because that is genuinely faster. The version of this story where somebody claims they never open a normal application again is not true, and would not be much use to you if it were.

Consolidating is the boring bit that makes the rest possible

Getting everything into one place sounds like a tidiness exercise. It is the thing the whole idea rests on.

You cannot untether yourself from ten applications, because there is nothing to talk to. Each one has its own front door, its own login, its own idea of what a record looks like, and the only thing joining them together is you, in the middle, moving information between them by hand.

Put the work in one place and there is suddenly a single thing to give an instruction to. Something that knows where the client files are, what is open, what was agreed last week, and can act on all of it in one go.

So the order matters. Consolidate first, untether second. Doing it the other way round gets you a voice assistant that can set a timer.

Away from the desk

Once everything lives in one place, the obvious question is why I need to be in front of it.

So this week I built a small thing that reads those voice notes. Not a polished dictation, just a brief, the way you would tell a colleague something on the way out of a meeting.

"Note to Claude, draft an email to the client at the advertising agency, letting him know we have not forgotten about him, ask how he is, keeping in touch really."

It works out that this is an email rather than an idea or a question, finds who I meant, reads that client's actual context, what is open and what we last discussed, and leaves a draft ready. When I next sit down I read it, change a word, send it.

Same with a question. "Note to Claude, investigate what this new design tool does and whether it touches our data." That comes back as a written answer with sources.

The prefix is deliberately a switch, so most of my voice notes stay ordinary voice notes and only the ones addressed to it become instructions. It never sends anything and never writes to a client's records on its own, because a voice note is a rough transcript of somebody talking near a road, and if it mishears me I want the damage to be a file I can delete.

The moments this recovers are entirely ordinary. The morning walk, where most of my decent ideas turn up anyway. Sat at a table. Waiting in a queue. Traffic. None of them long enough to open a laptop for, and collectively a large part of the week.

Why this keeps getting easier

The part I did not expect is the rate of change.

Nearly every week something lands that widens what is possible. Agents that can carry out a multi step job without me steering each step, that go and find what they need rather than waiting to be handed it, and that can now be pointed at a task and left to it. Something I worked around a month ago has stopped needing a workaround.

The effect is a change in what my job actually is.

Less and less of my time goes on building the thing from scratch. More and more goes on checking it, verifying it, and tweaking it. Reading what came back, deciding whether it is right, correcting the part that is not.

That is a different skill from the one I have spent years getting good at, and I do not think it is a downgrade. Judgement is harder to acquire than syntax. But the value is clearly moving, and if you build things for a living it is worth noticing which direction.

It also explains why working away from the desk matters more than it first looks. If the job were still mostly construction, being away from a keyboard would be fatal. If the job is increasingly instruction and review, the construction can happen while you are elsewhere, and only the judgement needs you.

What this does not do

It does not mean working all the time. That is the version of this idea I want no part of, and it is the obvious way to get it wrong.

The point is not that every gap in the day becomes billable. It is that a thought you have on a walk does not evaporate before you get back, so you are not spending the first hour at your desk reconstructing something you already worked out. That is less work, not more.

The review step also still lands on a screen. I read everything before it goes anywhere. Getting the checking off the laptop as well is a harder problem and I have not solved it.

What you can take from this

Consolidate before you automate. If your work is spread across a dozen tools, nothing can help you much, because nothing can see the whole picture. Getting it into one place is unglamorous and it is the prerequisite.

Try dictation before you try anything clever. Wispr Flow took ten minutes to set up and removed more friction than most of the automation I have built. Start there.

Give your capture a switch. Some of what you record is an instruction and most of it is not. Two words at the start of a sentence separates them.

Let rough input stay rough. You are handing over a brief, not dictating a finished document. Expecting your voice to produce polished work is what makes voice capture feel like a gimmick.

I am early into this and I expect to be wrong about parts of it. But the direction looks clear enough. The constraint was never how capable the tool is. It was that using it required a desk, and that is the part now changing.