Speak the task, not every punctuation mark
A coding agent is better served by a clear request than a stream of dictated syntax. Tell it what should change, how to tell the change works, and what it must preserve. For example: 'In the settings screen, make the microphone warning explain how to grant access. Keep the current layout and add a test for denied permission.' Review the transcript to catch a misheard file name before sending.
If you are editing code directly, voice recognition may struggle with symbols, case, and exact identifiers. Use the keyboard for a precise line or paste a verified symbol into the prompt. Treat voice as a fast way to explain the problem, not as a reason to skip exactness.
A four-part prompt that survives transcription
Make each spoken request short enough to inspect. A prompt that sounds natural but omits the acceptance condition can produce a convincing change with the wrong behavior.
- Name the location: the screen, function, file, or workflow the agent should inspect. Add a file reference in the coding tool when possible.
- Describe the desired behavior with a concrete before-and-after example.
- State constraints: preserve a public API, avoid a network call, or keep a privacy boundary intact.
- Ask the agent to show the diff and relevant test result. Review both before accepting or committing the change.
When the agent asks a follow-up, answer that one question. Long dictated monologues create more places for a transcription error or an unstated assumption to hide.
Pick the input path that fits your editor
VS Code's chat accepts natural-language prompts and file or symbol context. It also has an experimental Voice Mode. Anthropic documents a built-in /voice mode for Claude Code. If one of those already fits your workflow, compare it with ordinary Mac Dictation before adding another tool. A system-wide dictation app is useful when you move among an editor, terminal-backed agent, issue tracker, and messages and want the same input method in each text field.
The privacy boundary changes at Send. Local transcription can keep microphone audio and the draft prompt on your Mac, while a connected coding agent may receive the text, referenced code, and context you attach. Check the agent's settings and your project policy before including secrets or customer data.
Measure whether it actually helps
Try three real tasks: a bug report, a small feature, and a code-review comment. Compare time to a correct, reviewed result, including prompt repairs and code fixes. If an editor's built-in voice input wins, use it. If voice helps you explain intent but is awkward for exact symbols, keep the keyboard for the last mile.
Questions people ask
- Can I use voice with a coding agent?
- Yes, when its prompt accepts text. Dictate into that input, check names and constraints, then send. The agent's own voice features may be another option.
- Should I dictate source code line by line?
- Usually not for symbol-heavy edits. Dictate the task and expected behavior, then review or type exact identifiers and syntax where precision matters.
- Does local dictation keep my coding prompt private?
- It can keep speech processing local, but a connected coding agent may receive the prompt and context after you submit it. Those are two separate data paths.