Dictating with F8

The basics

• Click into any text field — email, doc, chat, terminal, code editor
• Hold F8. The pill lights up; a sound confirms recording
• Speak at a normal pace, including pauses and self-corrections
• Release (holds over 300ms finish on release). Your selected transcription engine recognizes speech, then the selected cleanup mode runs before paste

you say:   "send priya the invoice tuesday... actually wednesday morning"
you get:   "Send Priya the invoice Wednesday morning."

Hands-free mode

Double-tap F8 to start recording without holding. A quick tap (300ms or less) waits 380ms for a second tap; two taps within 400ms lock into hands-free. The pill shows elapsed state with confirm and cancel actions — press F8 again or ✓ to confirm and paste, ✕ to discard. Distinct sounds mark hands-free start, confirm, and cancel so you never wonder whether you are live.

Cleanup modes

• Verbatim: skips the second-pass cleanup model and dictionary replacement. Your selected speech-recognition engine still transcribes the recording.
• Light: cleans grammar and punctuation, fillers, stutters, and self-corrections. It can format clearly requested code, email-like content, and lists, but does not use your custom prompts.
• Enhanced: combines your global standing instructions with instructions for the active app. Use it for custom rewriting, formatting, and expansion.

App-aware handling covers code editors, browsers, chat apps, notes, mail, terminals, and other destinations. Light follows clear format requests; Enhanced applies the extra instructions you configured for the destination app.

Transcription & polish

Speech-to-text runs on your assigned engine. You can use on-device Whisper (download a model in the app) or a cloud speech provider such as Groq, OpenAI, or Gemini. Available models can change; use the current choices in Models & Providers. Before supported cloud transcription, Flunkey can include relevant Dictionary words in the recognition prompt so names like “Groq, Flunkey, VS Code” come out right.

Verbatim skips the second-pass model and dictionary replacement. Light applies cleanup without custom prompts. Enhanced combines global instructions with active-app instructions. The Dictionary F8 toggle controls dictionary behavior where supported; dictionary replacement does not run in Verbatim. Applicable second-pass cleanup uses the model assigned to that task. On-device Whisper stages audio and text in temporary files, deletes them after success or failure, and cleans crash leftovers on a later launch. If refusal detection spots a model lecturing instead of transcribing (“I can’t help with…”), nothing pastes and you get a Transcription blocked toast.

How paste works

Flunkey pastes via Ctrl+V (or Ctrl+Shift+V where the app profile calls for it) with per-app handling for code, chat, notes, terminals, and mail — then restores your previous clipboard contents when Preserve Clipboard is ON (the default). Every dictation is saved to Transcripts first, tagged with the app name, so a failed paste never loses text — just Ctrl+V manually. Empty results produce a Didn’t catch that toast instead of pasting silence.

Transcripts history

The Transcripts page keeps every dictation with its app name and date — entries show a 280-char clamp with expand + copy, and search covers both app and date. It is the fastest way to recover text when a paste misfires.

Dictation patterns

• Correct out loud: “at five… actually six pm” resolves to 6 pm
• Format by intent: “numbered list: first… second…” becomes a list
• Names stick: add people, products, and acronyms once in Dictionary & Polish
• Prompts count: choose Verbatim for long AI prompts when you want to avoid cleanup and dictionary replacement