transcribe

name	transcribe
description	Transcribe video and audio files via Gemini API. Use when the user asks to transcribe a recording, generate a meeting summary, extract speech from video or audio, or convert speech to text. Supports mp4, mkv, webm, avi, mov, mp3, wav, ogg, m4a, flac.

transcribe — video and audio transcription

Transcribes audio and video files via the Gemini 2.5 Flash API.

Modes

Generic (default)

Verbatim speech transcription with timecodes [MM:SS] and speaker identification.

Output files:

• <name> - transcript.md — verbatim speech with timecodes
• <name> - summary.md — short summary (with the --with-summary flag)

UI analysis (--analyze-ui, video only)

Detailed video analysis with breakdown of on-screen interface, navigation, and actions, plus screenshots.

Output files:

• <name> - summary.md — short overview (topic, participants, decisions)
• <name> - detailed.md — step-by-step chronological analysis with screen content and screenshots
• <name> - transcript.md — verbatim speech with timecodes
• screenshots/ — PNG frames of key moments

Usage

/transcribe <FilePath> [--output-dir DIR] [--analyze-ui] [--with-summary] [--format md|txt]

Parameter	Required	Default	Description
FilePath	yes	—	Path to audio / video file
--output-dir	no	<file_folder>/Transcript/<name>/	Result directory
--analyze-ui	no	off	UI-analysis mode (video only)
--with-summary	no	off	Add a summary (for generic mode)
--format	no	md	Output format: md or txt

Supported formats

• Video: mp4, mkv, webm, avi, mov
• Audio: mp3, wav, ogg, m4a, flac, aac, wma

Dependencies

• Python packages: google-genai, python-dotenv
• System: ffmpeg, ffprobe in PATH
• API key: ~/.cursor/skills/transcribe/.env (or your tool's skill folder) with GEMINI_API_KEY=...

Procedure

1. Determine FilePath and optional flags from the user's arguments.

2. Run the script:

PYTHONUNBUFFERED=1 python <skill-dir>/scripts/transcribe.py "<FilePath>" [--output-dir "<OutputDir>"] [--analyze-ui] [--with-summary] [--format md|txt]

IMPORTANT: PYTHONUNBUFFERED=1 is mandatory; otherwise stdout is buffered and progress is not displayed.

3. The script runs long (5–15 minutes depending on length). Files > 1 hour are split automatically.

4. After completion, report the result paths to the user.

5. Read the summary (if any) or the start of the transcript and show it to the user.

Cost

~$0.10 per 1 hour of recording (Gemini 2.5 Flash). Long files cost proportionally.

Limitations

• Maximum ~1 hour per Gemini request (the script splits automatically)
• Cyrillic file names: the script automatically copies the file to a temp location with an ASCII name
• Screenshot quality depends on source video quality
• Timecode accuracy +/- a few seconds
• --analyze-ui with an audio file automatically falls back to generic + summary