Voice-Driven Development: Speaking to Machines

The Bottleneck Is the Keyboard

Typing good prompts is slow. Not the physical act of typing (most engineers can type fast enough) but the cognitive overhead of composing structured, context-rich instructions while simultaneously thinking through the problem.

Speaking is different. When I talk through a problem, I naturally give more context, explain my reasoning, and surface assumptions I wouldn’t bother typing out. The output comes out richer because the medium doesn’t punish verbosity the way a keyboard does.

I’ve been using Superwhisper, an AI-powered voice-to-text tool for macOS, as a regular part of my workflow for almost a year now. Not just for AI prompts, but for email, documentation, notes, and general writing. But the biggest impact has been on how I interact with AI agents.

Why AI-Powered Dictation, Not Just Dictation

Standard voice-to-text transcribes exactly what you say. That’s the problem. When I speak naturally, I correct myself mid-sentence. I meander while I think. I rubber-duck, talking through an approach only to realise halfway through that it won’t work, then pivoting. A raw transcript of that is a mess.

Superwhisper processes dictation through an LLM before outputting text. It cleans up the natural speech patterns (the false starts, the corrections, the thinking-out-loud) and outputs a clean version that captures what I meant, not what I literally said. It handles file names and technical terms with correct formatting: camelCase, SCREAMING_SNAKE_CASE, kebab-case. When I rattle off a list of things conversationally, it structures them as an actual list. When I change my mind partway through a sentence, it takes the final version, not the journey.

The result is that I can talk to my computer the way I’d talk to a colleague, naturally, with all the messiness that implies. What comes out the other end is a well-structured prompt, email, or document.

Better Prompts, Not Just Faster Ones

The bigger benefit isn’t the speed gain, though that’s real, it’s that the prompts are better. When I type, I unconsciously economise. I might leave out context that feels obvious to me but isn’t obvious to the agent. I skip the reasoning behind a decision because explaining it feels like too much effort. When I speak, that friction disappears. I naturally say things like “the reason I want it this way is…” or “the constraint here is that…”. That kind of context makes the agent’s output significantly more useful.

This matters because, as I’ve written elsewhere, the spec is key to the quality of the product. The better your instructions, the better the output. Voice dictation removes a barrier between what I’m thinking and what the agent receives.

The Private Office Advantage

There’s an obvious practical constraint, you need to be able to talk to your computer. Working from a private home office, this isn’t a problem. I don’t know how well this would work in a shared office space. Headsets and noise-cancelling can help, but there’s a social awkwardness to dictating prompts while colleagues are trying to concentrate three feet away. For now, this is an advantage of remote and private working environments.

Where This Is Going

Voice dictation feels like an interim step in a longer arc of human-machine interaction. The fundamental problem is bandwidth. Getting what’s in my head into the machine with as little loss and friction as possible. A keyboard is a narrow, lossy channel. Voice is wider. But it’s still indirect. I’m encoding thoughts into speech, which gets decoded back into text, which gets interpreted by an agent.

The obvious question is: what comes next? Interfaces that can infer intent more directly? Input mechanisms we haven’t imagined yet? The history of computing is a history of narrowing the gap between thought and action, from punch cards to command lines to GUIs to touch to voice. Each step removed a layer of translation.

What’s already clear is that AI agents aren’t just the next step on that trajectory. They represent a fundamentally different kind of interaction. Every previous shift, from punch cards to GUIs to voice, changed how we input instructions. The relationship stayed the same: you tell the machine what to do, and it does it. With agents, you describe intent, and the machine determines how to achieve it. That’s not a new input mechanism. It’s a new interaction model entirely, one where the machine is a collaborator rather than an executor. We’re still in the early stages of understanding what that means for how we work.