Treat the input as a URL or local media path; no action keyword is needed. Infer the input from natural-language requests, or ask if it is missing.

Transcript rules

Prefer full-length text subtitles. For URLs, prefer uploaded subtitles over original auto-captions. Use speech recognition only when no suitable subtitles exist; failed inspection, download, extraction, or cleanup means stop and report the error, not fall back. Ask when track choice, language, or coverage is unclear, including when only translated or wrong-language subtitles exist. Do not install tools without permission.

Default subtitle selection to English unless the user specifies a language. For recognition, pass the spoken language only when known; otherwise let Whisper detect it. Transcription does not translate into the requested subtitle language.

Setup and routing

Save outputs under <cwd>/.ai/transcript unless the user specifies another folder. Create a new directory for each media item; never reuse an existing one. Use its absolute path as <run-dir> in the commands below. Process playlist items and local files separately.

• URL: follow URL transcripts. If this returns subtitles, clean them below. If it returns downloaded media, go directly to Speech recognition, skipping local subtitle inspection.
• Local path: follow Local transcripts to find subtitles or recognize speech.

Cleanup and output

Run this skill's cleanup script on the selected subtitle:

node "<skill-dir>/scripts/strip-transcript.mjs" "<subtitle-file>" > "<run-dir>/transcript.txt"

Do not use --in-place: it deletes the source and derives a different output name. Add --dedup only for rolling auto-captions that repeat text across cues. Use Whisper's TXT directly without cleanup.

Check command errors and confirm this run produced a transcript containing text. Report empty results rather than claiming success. After successful subtitle cleanup, delete only downloaded or extracted subtitles inside <run-dir>, unless the user wants to retain them. Never modify or delete the user's media or sidecars; retain downloaded media used for recognition.

Show <run-dir>/transcript.txt, identify subtitles or speech recognition as its source, and offer to summarize it. For explicitly requested raw subtitles or structured output, return that format instead.