How to dictate code and AI prompts in Cursor on a Mac

Start with tasks where spoken context is more valuable than exact punctuation. Cursor's Agent can work from a detailed natural-language request, and its official documentation describes Agent as able to search, edit, and run commands while preserving a reviewable diff. Dictate the goal, files or modules in scope, current behavior, expected behavior, constraints, tests, and explicit non-goals. That is usually faster and safer than speaking the final implementation.

Other good targets are Ask-mode questions, code-review notes, commit-message drafts, issue reproduction steps, README sections, docstrings, comments, test names, and short identifiers. Direct code dictation is most useful for small fragments where a spoken symbol vocabulary is predictable. If you find yourself narrating indentation and every delimiter, stop and switch to autocomplete, an agent instruction, or the keyboard.

Choose a local transcription model first so the baseline can run without uploading source-related speech. Create a Cursor-specific mode and enable code dictation only for the mode used in editor-like fields. Add a small project vocabulary containing repository names, frameworks, domain terms, and frequently used identifiers. Prefer deterministic replacements for terms that must always have one spelling.

Set a global push-to-talk or toggle shortcut that does not conflict with Cursor's commands. Cursor uses VS Code keybindings as a baseline and its official shortcut reference assigns important combinations to Agent, chat, inline edit, terminal prompts, and the command palette. Test the chosen shortcut in the editor, chat, terminal, source-control message, and a modal. A collision can silently trigger an editor command while you think you are recording.

Use a fixed spoken structure: objective, context, constraints, verification, and stop condition. For example, say that you want a failing checkout test repaired; identify the relevant route and existing behavior; prohibit schema or dependency changes; ask for the narrowest patch; require the targeted test and type check; and instruct the agent to stop if credentials or a product decision are missing. This structure reduces ambiguity without requiring a perfectly written paragraph.

Pause before submitting. Read the transcript for missing negation, wrong file names, changed numbers, or a phrase such as “do not” that cleanup removed. Cursor's chat input has its own send and queue shortcuts, so keep dictation delivery separate from message submission. A system-wide tool should paste text; you should decide when Agent receives it.

Use short semantic units. Dictate an identifier with a casing command, inspect it, then continue. Speak punctuation only when the mapping is unambiguous. Keep comments and string content in a prose mode so words like “dot” or “colon” are not transformed unexpectedly. When switching between chat and source, verify the active mode instead of relying solely on app detection because both fields live in Cursor.

Immediately run the formatter and language diagnostics. A formatter can normalize spaces and indentation, but it cannot know whether the wrong operator or identifier was spoken. Type checks catch some category errors; tests catch expected behavior; diff review catches changes neither tool understands. Dictation can accelerate input while the normal engineering system remains responsible for correctness.

A local model reduces external exposure, but system-wide dictation still uses sensitive permissions. Apple requires explicit microphone permission, and an app that inserts text through Accessibility APIs receives broad control. Grant access only to an app you trust, review it in System Settings, and revoke it when no longer used. Confirm whether audio, transcript history, crash reports, clipboard contents, backups, or optional enhancement services retain data.

Never speak secrets into source or chat. Recognition history can preserve them, the target app can index them, and an AI agent can include them in context. Use secret managers and placeholders. Keep terminal auto-run disabled or confirmation-gated for dictated prompts, and inspect the resolved command and target path before execution. The most dangerous voice error is not a typo; it is a plausible destructive instruction sent without a pause.