Hold a hotkey, speak, and your words become an engineering prompt — transcribed locally, refined by an LLM, injected into your AI dev tool. Built entirely with AI.
🎙️ Push-to-Talk Local transcription · LLM refinement · One-tap injection
Screenshots and a demo video are coming soon. The chain below is already working end-to-end.
Hold the global hotkey (default Right Option) and speak naturally. A 200 ms hold-to-confirm guard prevents mistouches.
Release to transcribe on-device with WhisperKit small — audio stays in memory and is never written to disk.
Your own LLM key turns the transcript into a clean engineering prompt, with hard anti-hallucination constraints.
The result lands at the cursor of the app you were using — clipboard + Cmd+V with restore and race protection.
This project was entirely built with AI coding assistants. Every line of code was generated, reviewed, and refined through AI collaboration.
Kimi K3
V4 Pro
Default Right Option with 200 ms confirmation, 300 ms mistouch cancel and Esc to discard. Presets: Right Option / Right Command / Right Shift / Fn.
WhisperKit small runs fully on-device with Chinese/English mixed support; Speech framework fallback until the model is ready.
Removes filler words, normalizes phrasing, resolves anaphora from context, and falls back to raw text on timeout — bring your own key.
Snapshots the frontmost app and window title on press, re-checks on release, and classifies terminal / editor / browser / other.
Clipboard snapshot/restore with change-count race protection; optional auto-send (Return); newline folding for CLI targets.
Waveform, stage status and success/failure feedback — visible in full screen and across Spaces without stealing focus.
Guided onboarding with deep links into System Settings and graceful degradation for mic / input monitoring / accessibility.
Categorized os_log plus on-disk logs with one-click export from Settings.