Transcribe Audio and Generate Meeting Minutes

This skill automates audio-to-text transcription into professional Markdown output, extracting technical metadata (speakers, timestamps, language, file size, duration) and generating structured meeting minutes and executive summaries. It uses Faster-Whisper or original Whisper with zero configuration, working across projects without hardcoded paths or API keys, and is platform-agnostic - it runs in any terminal context where GitHub Copilot CLI is available, without depending on specific project configuration or external APIs.

The workflow starts with tool discovery: it checks for Faster-Whisper first (preferred, described as 4-5x faster than the original), falls back to standard Whisper, and checks separately for ffmpeg to enable format conversion. If no transcriber is found, it offers one-confirmation automatic installation via a bundled install script, falling back to manual pip/brew instructions if that script is missing or fails. Audio file validation accepts a local path or a URL (downloaded to a temp directory), confirms the file exists, extracts file size and duration/codec via ffprobe, warns when a file exceeds 25MB since processing may take several minutes, and converts unsupported formats to WAV via ffmpeg when possible - supported formats are MP3, WAV, M4A, OGG, FLAC, WEBM, and MP4.

Output is a Markdown report with a metadata table (filename, size, duration, language, speaker count, transcription engine and model), meeting minutes (participants, topics discussed with timestamps and key points, decisions made, and action items with assignee and due date when mentioned), generated via Python helpers that cluster transcript segments by topic and detect decision/action-item keywords, plus an AI-generated executive summary using a Chain-of-Density-style prompt capped at a configurable number of paragraphs. Output files are timestamped to avoid overwriting (transcript-{timestamp}.md and, when LLM processing is used, ata-{timestamp}.md for the processed meeting-minutes version), with temporary metadata/transcription JSON files cleaned up afterward.

When a user supplies a custom summarization prompt, the skill can optionally improve it automatically (invoking a prompt-engineer helper via gh copilot), show both the original and improved versions side by side, and let the user pick which to use before running it through an LLM CLI tool (Claude or GitHub Copilot) with a progress spinner and a 5-minute timeout.

Invoke this when a user needs to transcribe audio or video to text, wants meeting minutes automatically generated from a recording, needs speaker identification, needs SRT/VTT subtitle formats, wants an executive summary of long audio, or has audio files in one of the supported formats. It fits meetings, interviews, lectures, and general content analysis; batch processing of multiple files (for example an entire recordings folder) is also supported, processing each file sequentially and reporting per-file completion times.

A representative metadata-extraction step:

# Get file size
FILE_SIZE=$(du -h "$AUDIO_FILE" | cut -f1)

# Get duration and format using ffprobe
DURATION=$(ffprobe -v error -show_entries format=duration \
    -of default=noprint_wrappers=1:nokey=1 "$AUDIO_FILE" 2>/dev/null)
FORMAT=$(ffprobe -v error -select_streams a:0 -show_entries \
    stream=codec_name -of default=noprint_wrappers=1:nokey=1 "$AUDIO_FILE" 2>/dev/null)

Input is an audio/video file path or URL; output is one or more timestamped Markdown files (a raw transcript and, optionally, an LLM-processed meeting-minutes document), plus optional SRT/VTT subtitle and JSON structured-data exports.

Anyone needing quick, structured documentation from recorded audio - meeting facilitators wanting automatic minutes and action items, researchers transcribing interviews, or teams processing a batch of recordings without setting up API keys or project-specific configuration.