Brian Mills

Case study · A tool I built · Personal project

Purr: a tool for talking instead of typing.

Three

versions of everything I say, produced on the spot: one raw, one clean, one ready to hand to an AI. The audience decides which.

The problem

I think faster than I type, and a lot of my work starts as me talking through an idea. Ordinary dictation does not help much, because it gives you one block of messy text full of false starts and filler, and then you have to clean it up by hand anyway. The bottleneck moves, but it does not go away.

What I built

Purr is a small voice-to-text tool I wrote for my own Windows setup. A hotkey opens a little panel over whatever app I am in, usually the Claude desktop app. I talk, and it stops on its own when I go quiet. The audio is transcribed locally on my own GPU with a speech-to-text model, so nothing leaves the machine. A local language model then rewrites that transcript three ways, each with its own set of instructions: one to preserve it word for word, one to clean it up, one to structure it for a machine. I pick one, and it drops into the app I was already using.

The part that is actually about communication

The three versions are the reason Purr is on this site. They are the same words written for three different uses, which is the same audience-segmentation work I do everywhere else, turned on my own voice. Take one thing I might say out loud:

"um so the main thing is we we need to get the newsletter out by friday i think, and somebody should double check the outlook version before it goes"

Raw

For the record. Nothing removed.

um so the main thing is we we need to get the newsletter out by friday i think, and somebody should double check the outlook version before it goes

Clean

For a person to read.

The main thing: we need the newsletter out by Friday, and someone should check the Outlook version before it sends.

Prompt

For an AI to act on.

Task: send the newsletter by Friday. Before sending, test the Outlook rendering. Flag any blockers to the timeline.

Same sentence, three readers. The raw version keeps the exact record. The clean version is what a colleague should see. The prompt version is structured for a machine that needs a clear instruction, not a paragraph. Deciding what each reader needs, and giving them exactly that, is the work. Purr just made me do it to myself.

I talk transcribe local, on GPU local model three versions RawCleanPrompt record people machines
Voice in, three audience-tailored versions out. Illustration built for this portfolio.

Where it stands

Purr is a working personal tool, not a product. I use it daily, including to draft early versions of the writing on this site. It is the clearest proof I can offer that the receiver-first method is not just how I write at work. It is how I think.

Next

Why I write anything three ways before I send it once: the method is under Writing.