// the find
yetone/voice-input-src
This repo is a README and a pointer to yetone/voice-input-dist, and the README is a single Claude Code prompt asking for a Swift menu-bar app. The app holds the Fn key to dictate through Apple Speech, optionally runs an OpenAI-compatible LLM cleanup pass, and pastes the result into the focused field. It is aimed at macOS users who mix Chinese and English while dictating and want technical terms corrected.
Specifying a CGEvent tap that suppresses the Fn event targets the first thing most Fn-key tools get wrong, which is the emoji picker firing on the same press. Switching to an ASCII input source before Cmd+V and restoring the previous one addresses a real failure where a CJK input method swallows the paste, and restoring the clipboard afterward is specified too. The LLM correction prompt is limited to homophone and English-term fixes and says to return correct input unchanged, which is the right bias for a cleanup pass that should not rewrite the user's words.
The README is a generation prompt, not documentation. It does not say what the repo contains, how someone installs the built app, or how the dist repo relates to this one, and the visible tree has no Swift sources. The event tap and paste injection need Accessibility and Input Monitoring permissions on macOS, which the spec never mentions, so a first run will likely look broken to anyone who has not granted them. The LLM step blocks injection with no timeout or fallback, so a slow or unreachable endpoint leaves the user staring at 'Refining...' with no way to get the raw transcript.