// Agents and skills · Tech //

Data · runs in Claude Code

It takes the project's audio or video (and, if you want, the script, from which it pulls the proper names for Whisper's prompt). It outputs `transcricao/palavras.json` with the start and end time of every word and its confidence, plus `segmentos.json` for reading, `texto.txt`, and a `qa.json` with low-confidence words, gaps with no speech, and the log of how the chunks were stitched together. It trims the silence at the start before transcribing (that's what keeps captions from drifting ahead in long videos), fine-tunes the start of each word using the audio energy, and also searches for terms, phrases, or misspelled names inside the transcript, with the minute. For people who edit video and need captions, timeline, motion graphics, and sound effects all anchored to the same timing as the original media.

• Data
• Runs in Claude Code
• Intermediate