Voice to text for Claude Code, Cursor and every other AI agent

The actual problem

Short prompts are a typing artefact, not a choice

When you type a prompt, you unconsciously minimise. Every clause costs keystrokes, so the thought that started as “refactor this module, here is why the current shape is wrong, here is what I want instead, watch out for the migration path, and do not touch the auth code” arrives at the agent as “refactor this module.” The agent answers the prompt you sent, not the one you meant, and you spend the next three exchanges adding back the context you edited out to save your fingers.

Speaking does not have that tax. The full version takes no more effort to say than the short one, so you say the full version — and the agent gets the picture the first time. Developers who have made this switch consistently report that the payoff is fewer rounds rather than faster words, which is a different and much larger saving than it first sounds.

The arithmetic

Eighty seconds against five minutes

People speak at around 150 words a minute and type at around 40. Run that over a realistic feature spec and the gap is not subtle.

Straight arithmetic from typical speaking and typing rates — check it with a calculator. It deliberately excludes thinking time, which is the same either way, and it is the smaller half of the benefit.

Where it pays, and where it does not

The same shortcut works in all of them

You do not need a plugin or an integration for any of these. Each one treats its prompt box as an ordinary text field, so a system-wide dictation shortcut simply works — the only thing to check is that its hotkey does not collide with one your editor already uses.

One workflow note that costs nothing and saves a lot: read the prompt before you submit it. An agent commits to whatever you actually said, so a spoken paragraph with two contradictory sentences produces a change with two contradictory halves. Dictation raises your throughput; it does not lower the bar for saying what you mean.

The detail that decides it

“Cart” keeps coming out as “card”

Ask developers who tried voice input and went back, and this is the reason. Generic transcription has no idea what your codebase is called. It writes card for cart, mangles library names, splits identifiers, and turns your project's vocabulary into near-misses. Every one of those is a correction you have to type, which erases the saving you came for — and it happens most on exactly the words that carry the meaning.

Dictro learns the words you correct. Fix an identifier once and it stops coming back wrong, so your project's nouns survive the trip from your mouth to the prompt box. The same mechanism covers library names, service names, and the jargon that only exists inside your team.

It is also worth knowing where the audio goes. Prompts contain your architecture, your table names, and sometimes your customers. Dictro's recognition and cleanup both run on your own Mac, so none of that leaves the machine — see how on-device dictation works.

Be honest about the edge

You should still type the code

Dictating actual source is a bad idea and the people who do this every day agree. Symbols, brackets, precise identifiers, a gnarly regex, an exact SQL fragment — all faster and more reliable from the keyboard. The split that works is: speak the English, type the syntax.

That split turns out to cover most of the day once AI agents are doing the typing of code. Prompts, specs, commit messages, pull request descriptions, review comments, the long “here is what I am actually trying to do” context dumps — that is English, and it is where the hours go.

Dictro runs on Apple Silicon Macs on macOS 15 or later, and on Windows 10 and 11. Recognition and cleanup run on your machine; no audio or transcript is uploaded. Free plan: 10,000 words a week, no card.

For dictating inside the editor rather than into an agent, see dictation for developers. If your hands are the reason you are here, read dictation for RSI. Or start on the free plan.