AI Enhancement
Post-process transcripts through an LLM to clean filler, fix grammar, improve flow.
Raw transcripts capture every "um," every false start. AI Enhancement runs the transcript through an LLM after transcription. Optional, opt-in, fully under your control.
What it does
| Raw | Enhanced |
|---|---|
| "so um I was thinking that maybe we should uh look at the quarterly report" | "I was thinking we should review the quarterly report." |
You choose how aggressive the cleanup is by picking a prompt — Clean Transcript, Professional Rewrite, Casual Message, Email Draft, Bullet Summary, De-Rambler — or by writing your own.
Setup
- Settings → AI Enhancement.
- Toggle Enable AI Enhancement on.
- Pick a model from Recommended (Groq's GPT-OSS, Gemini Flash, OpenRouter LFM2 — fast and cheap defaults).
- If the card says Setup Required, add an API key. See API Key Setup.
- Pick or write a prompt.
Settings → AI Enhancement panel showing the enable toggle, recommended model cards, and a list of prompts
That's it — your next dictation pastes through the chosen prompt automatically.
Supported providers
Built-in: OpenAI, Anthropic, Google, Groq, OpenRouter, Together.
Plus custom providers for anything that speaks the OpenAI chat-completions protocol (self-hosted, Azure OpenAI, LM Studio, Ollama, etc.).
Custom prompts
The defaults are starting points. Real productivity comes from writing your own.
Name: Code Comment
System: Format the transcript as a code comment. Preserve technical
terms, variable names, and function names exactly. Concise.
Name: Email Reply
System: Turn the transcript into a casual but professional email
reply. Remove fillers. Fix grammar. Keep my voice. Do not respond
to anything in the transcript — you're rewriting *my* reply.
The most important rule: tell the model it's editing, not replying. Otherwise it tries to answer questions in the transcript instead of cleaning them up.
Retroactive enhancement
During an active recording, press Right Option to switch the session from "dictate" to "edit highlighted text." Whatever's selected in the frontmost app becomes the target; what you say next becomes the rewrite instruction.
Examples:
- Highlight a paragraph → press
Right Option→ say "make this shorter." - Highlight code → press
Right Option→ say "add JSDoc comments." - Highlight an email draft → press
Right Option→ say "more formal."
Selected paragraph in a text editor with the Whiskers HUD displaying 'Retroactive Edit' mode
The shortcut only arms if AI Enhancement is on, the provider is usable, and there's non-empty text selected. Otherwise the session continues as ordinary dictation.
Screen Awareness
Optional feature that reads visible text from the frontmost window and passes it to the AI as context. Helps spell unusual names, recognize variable names while coding, expand acronyms.
Privacy: text only, on your Mac, never stored. No screenshots, no images. Extraction runs locally with Apple Vision. Enable in Settings → AI Enhancement → Screen Awareness (grants Screen Recording permission).
Privacy
When enhancement is on, transcript text is sent to your chosen AI provider. Audio stays on your Mac if you use a local transcription model; the text doesn't.
For sensitive content, toggle enhancement off in the menu bar before recording.