Stop Typing
I do not type prompts anymore. I talk to my computer. The difference is not convenience. It is a fundamentally different relationship with the work.
I used to type everything. Every prompt, every plan, every correction. Carefully composed, carefully punctuated, carefully wrong. Because typing makes you edit before you think. You craft a sentence while the thought is still forming, and the sentence wins. The real idea — the messy, provisional, half-formed thing that might actually be right — gets smoothed away in the act of putting it into words on a keyboard.
So I stopped.
I talk to my computer now. Not through a voice assistant, not through dictation in the traditional sense. I use speech-to-text — MacWhisper, though the specific tool matters less than the practice — and I speak my thoughts as they arrive. Unstructured. Unpunctuated. Full of false starts and corrections and the kind of rambling that would embarrass me if anyone were listening.
The language model does not care. It understands perfectly. It does not need my prose to be clean. It needs my thinking to be honest, and honest thinking rarely arrives in well-formed paragraphs.
This changes more than speed, though the speed is considerable. A prompt I would spend three minutes typing takes thirty seconds to speak. Over a day of development that difference is not trivial. Over a week it is transformative. But speed is the lesser benefit.
The greater benefit is this: when you speak, you explain. When you type, you instruct. These are different acts. Instructions are compressed, imperative, stripped of context. Explanations are expansive, discursive, full of the reasoning behind the request. And language models — this should not surprise anyone, given what they are — perform dramatically better when they understand why you want something, not just what.
A typed prompt: “Refactor the authentication module to use JWT tokens.” A spoken prompt, transcribed: “So the authentication module is using session-based auth right now, which was fine when we had a single server, but we’re moving to a distributed setup and sessions don’t travel well across instances. I think JWT makes sense here because we can validate tokens without hitting a central store. But I’m worried about token revocation — we need to think about that. Can you refactor the auth module to use JWTs and flag anywhere that revocation would be a concern?”
The second prompt is three times longer and five times more useful. It contains the reasoning, the constraints, the concern about revocation that the developer knows matters but would have forgotten to type. The model now produces something informed rather than something merely obedient.
I do not compose my prompts. I think aloud. The model gets the raw material of my reasoning — the context, the caveats, the half-articulated worry about an edge case — and it produces better work because of it. Not because it is smarter. Because I have given it more to work with.
There is a resistance to this. It feels undignified. It feels unserious. Software engineers are people who type. We have mechanical keyboards with specific switches and particular keycap profiles and we are proud of our words-per-minute in a way that borders on the devotional. Speaking to a computer in a stream of consciousness feels like a regression, like abandoning a skill we spent years refining.
It is not a regression. It is an admission that the skill we refined was optimised for a different task. Typing is for authoring. Speaking is for thinking. And the work, as I have argued throughout this series, is no longer authoring. It is thinking — directing, evaluating, steering. Your voice is better suited to that work than your keyboard. Not because it is faster, though it is. Because it is closer to thought.
Try it for a day. Just one. Speak every prompt. Do not type a single instruction to the model. You will feel foolish for the first hour. By the afternoon you will not go back (unless you find yourself pairing with another human!).
This is the third in a series on LLM-assisted engineering practices. Previously: Your screen is the bottleneck and Generate less.
Next: why the most important thing you can build is not a feature — it is the system that tells you whether your features work.
Find me on LinkedIn or at [email protected].