Check Talkie’s speech setup
Open Talkie Settings › Models › Transcription. If the Vox model list is empty, start Vox and choose Refresh models. For an available model marked Not installed, click its row to download, then click the downloaded row to load and select it. Check Selected or Loaded and resolve loading errors. The reviewed default path uses Apple Speech for the live preview and Parakeet v3 through Vox for the final transcript after capture. Confirm both dependencies; a working preview alone does not establish that the final prompt text is ready. Speech setup is separate from your coding agent’s model.
Check the microphone and recording trigger
Open Settings › Dictation › Recording. Under Microphone, choose the input device you intend to use and check its live level while you speak. Under Triggers, check Toggle recording: press its configured shortcut once to start and again to stop. If you prefer holding a key, turn on Enable push-to-talk and check the Push-to-talk shortcut. Use the shortcut shown in your settings rather than assuming a default.
Choose Clipboard Only for a review step
Open Settings › Dictation › Delivery. Under Paste action, select Clipboard Only. After transcription, Talkie copies the text without pasting it. You can then focus the intended input and press ⌘V. Copy & Paste is the automatic-paste option; delivery can also be configured to press Enter after pasting. Start with Clipboard Only when you want to inspect each instruction before it reaches an agent.
Open the agent’s prompt field
For Claude Code, open a running Claude Code session and focus its prompt input in your terminal. For Codex, open the task you want to work on and focus its composer. For Cursor, open the agent chat and focus its input. A terminal running a coding agent and a shell waiting for a command are different destinations. Check which one has focus before pasting.
Speak the problem, constraints, and a useful next action
Use voice to explain behavior and intent. For example: “The settings panel loses my selection when I reopen it. Find where that selection is stored, explain the cause, and propose the smallest fix. Preserve the other settings. Add a regression test if the existing test setup can cover it.” Give the agent the symptom, the boundary, and what you want it to do next.
Review names and scope before sending
Stop the recording, paste the transcript into the prompt field, and read it. Correct file paths, symbol names, numbers, negations, and punctuation that changes meaning. Paste exact code or error output from its source instead of relying on speech to reproduce it. Send the prompt with the agent’s normal send action once it says what you intend. Dictating text does not replace the agent’s tool approvals or your review of its changes.
Use voice for the next turn, too
After the agent responds, dictate a focused follow-up: “Show me the callers before changing that interface,” or “Keep the fix inside this module and run its existing tests.” Attach the relevant diff, error, or screen when the agent needs evidence. A spoken description such as “the button on the left” does not tell an agent what is on your screen unless you also supply that context.
Keep the thought and its context available
Talkie also keeps captures beyond the destination text field. A saved dictation can carry its source app, window title, and a browser URL when available, which can help you place the thought when you return to it. History depends on your Dictation Retention setting in Storage. Save a longer voice memo or capture the relevant screen when the idea needs more context, then supply the material you want the agent to use. Capturing it in Talkie and attaching it to an agent are separate actions.
If recording or delivery does not work
If recording does not start, check the configured trigger and macOS Microphone permission for the Talkie copy you are running. If the recording is silent, check the selected input and its live level. If transcription is unavailable, check Vox and Models › Transcription; refresh an empty list and resolve installation or loading errors. With Clipboard Only, focus the prompt field and press ⌘V. With Copy & Paste, check field focus and Accessibility permission. If Talkie reports that a transcript was held, it is on the clipboard; return to the intended input and paste it there.
Separate local dictation from the agent’s data path
The reviewed live speech path uses on-device recognition through Vox. Prepare any required language or model assets before checking offline use. Once you send the prompt, the coding agent processes it under its own account, workspace, and provider settings. Optional Talkie AI transforms, sync, and meeting transcription have separate network paths, too. Local recognition does not make an entire coding session offline or keep material local after you send it to a remote service.