Cursor’s agent doesn’t want code from you. It wants a brief: what to build, where, the constraints, and what done looks like. That’s a paragraph of English, and English is what you’re good at saying out loud. This is a workflow that’s held up over a few months of daily use.
Where voice helps
The agent panel. This is 80% of it. Describing a change takes 20 seconds to say and two minutes to type with the same detail. Detail is what stops the agent guessing.
Commit messages and PR descriptions. You already know what you did. Say it.
Comments and docstrings. Explaining why in a sentence is easier spoken.
Where it doesn’t
Writing code itself. Symbols, indentation, exact identifiers: type them. Dictation is for the prose around the code, not the code.
What to say
The agent responds to structure. A brief that works has four parts, and you can say them in any order because Speaky Box’s Smart key arranges them:
- The task. “Add a retry to the upload function.”
- Where. “It’s in
api/upload.ts.” - Constraints. “Don’t change the function signature, keep the existing error type.”
- Done when. “Tests pass and a failed upload retries three times with backoff.”
Say that as one ramble (“ok add a retry to the upload thing in api upload, don’t change the signature, um, and keep the error type, done when it retries three times and the tests pass”) and what lands in the agent panel is a clean four-line brief.
Getting names right
The one thing that breaks voice coding is spelling. useEffect, pgvector, your internal service names. Add them to Speaky Box’s Words list once and they’re spelt correctly every time, in every app. Ten minutes of setup on day one saves a correction every hour after.
The habit
Hold the key, talk, let go, read the brief, hit run. If the brief is wrong, hold the key and say the correction. After a week you stop typing into the agent panel entirely, and you write better briefs, because saying “and don’t touch the migration” costs nothing.