What problem does it solve? Turning a raw voice or video recording into a usable written transcript requires more than speech recognition: the recognizer's output contains mis-hearings, filler, and tangles that need human-level judgement to clean without rewriting the speaker's voice. This Skill bridges that gap by combining a deterministic transcription script with editorial rules for producing a faithful, readable transcript file. ## Core Features & Use Cases - Deterministic transcription pipeline: Runs scripts/transcribe.py to extract audio, call Deepgram, and save a timecoded transcript plus the raw API response for later correction. - Three output modes: verbatim (subtitle-ready, speaker's exact words), retake (a cleaned script to re-record from, with footnoted edits), and prose (connected text for repo context). - Accountability for guesses: Every uncertain recognition is marked with a timecode and listed in a shrinking table so the operator can verify corrections against the recording. - Use Case: You record a six-minute video diary on your phone, commit it to the repo, and ask for a transcript. The Skill produces writing/<project>/dictations/<slug>.md with a summary, the cleaned transcript, marked guesses with timecodes, and an afterword — ready for subtitling or turning into a post. ## Quick Start Transcribe the attached recording into a dictation file using the dictation skill, in retake mode, with the slug "launch-thoughts".
