I Stopped Typing Prompts to My AI Coding Agent. I Talk to It Now.

There's a quiet workflow shift happening among developers who use AI coding agents: they've stopped typing their prompts. They talk.

It sounds like a gimmick until you notice the mismatch that makes it work. The best prompts to Claude Code, Cursor, or any coding agent are long, specific, and conversational — "look at the auth middleware, the session refresh is firing twice on tab focus, I think the listener isn't being cleaned up, check the useEffect in SessionProvider and fix the teardown." That's a forty-word prompt. Forty words is a chore to type. It's a ten-second sentence to say.

Typing pressure pushes you toward short prompts. Short prompts get worse results. Dictation removes the pressure, so you naturally give the agent the context it actually needs. The prompt quality goes up because the input cost went down.

Why this is suddenly everywhere

Agents got good enough that the bottleneck moved. A year ago the model was the constraint; now, for a lot of tasks, the constraint is how fast and how completely you can tell it what you want. Voice is simply a higher-bandwidth channel for intent. You think out loud at 150 words per minute; most developers type at 50-70.

There's a second, less obvious reason: prompts are not code. Code punishes imprecision — one wrong character breaks everything, which is why dictating code never took off. Prompts are the opposite. They reward the rambling specificity of natural speech. All the "and also," "wait, actually," "the thing I mean is" texture that would be terrible in a function body is exactly what makes a prompt rich.

My setup

I use IndianWhisper (disclosure: I build it) with a push-to-talk key. The flow:

1. Cursor or terminal focused, Claude Code waiting for input

2. Hold the push-to-talk key, describe what I want — as long as it needs to be

3. Release. The text lands at the cursor, cleaned up — fillers stripped, punctuation fixed

4. Enter. Agent goes to work

The details that matter for this workflow specifically:

It has to work offline and on-device. Your prompts contain your architecture, your bugs, your client's business logic. Piping that audio through a third-party cloud on top of the agent you're already trusting is a second data exposure most client contracts never contemplated. On-device transcription means the audio never leaves the machine — the only thing that goes anywhere is the text you were going to send the agent anyway. It has to handle how you actually speak. I think in Hinglish. Mid-prompt I'll say "iss function ko refactor karo but keep the return type same." A dictation engine that panics at code-switching turns that into soup. One built for it doesn't blink. The same applies to speaking technical vocabulary with an Indian accent — "kubernetes," "postgres," "middleware" should not come out mangled. Cleanup has to be automatic. Raw speech is full of "um, so, basically." A dictation tool with an LLM cleanup pass hands the agent tight text without you thinking about it.

Where it works beyond prompts

Once the habit forms, it spreads. Commit messages — spoken in five seconds while your hands leave the keyboard. PR descriptions. Slack updates explaining what you shipped. Code review comments, where tone matters and typing makes everyone terse. The rubber-duck effect is real too: describing a bug out loud to your agent is rubber-duck debugging where the duck can actually fix it.

Try the workflow

Any dictation tool can technically do this. If you want the on-device, code-switching-friendly version: IndianWhisper is free to download for Mac and Windows, with a Chrome extension for browser work. Set the push-to-talk key, open your agent, and describe your next bug out loud instead of typing it.

The first time you watch a long, precise prompt appear from a ten-second sentence — and the agent nails it on the first pass because you finally gave it enough context — the keyboard starts feeling like the slow way to talk to a machine.

— Dhruv