| name | transcription-workflow |
| description | Process audio/video files into structured content โ transcribe, clean up, extract key points, and organize into chapters. Use when a user uploads an audio/video file or asks to work with a transcript. |
Transcription Workflow
Trigger
User uploads an audio or video file (.mp3, .mp4, .webm, .wav, .m4a).
Steps
- Start transcription โ use
start_transcription tool with the uploaded file's S3 key
- Inform user โ tell them transcription is in progress (takes 1-5 min depending on length)
- When notified of completion โ use
list_pending_jobs to get the result
- Mark consumed โ use
mark_job_consumed after processing
- Read transcript โ use
read_file with the transcript S3 key
- Process โ based on user intent:
- Clean up raw transcript (remove filler words, fix formatting)
- Extract key topics and timestamps
- Create chapter markers
- Summarize main points
- Convert to video script outline
Transcription Output Format
The raw transcript is plain text (Hebrew or English). AWS Transcribe provides:
- Full text without timestamps (current implementation)
- Language auto-detection
Post-Processing Options
Ask the user what they want:
- ืกืืืื (Summary) โ bullet points of key topics
- ืชืืื ืืช ืกืจืืื (Video plan) โ convert talk into a YouTube video structure
- ื ืืงืื (Cleanup) โ readable version without stutters/fillers
- ืคืจืงืื (Chapters) โ segment into logical sections with titles
Important Notes
- Transcription is async โ don't block the conversation
- Files must be in the session's uploads folder
- After transcription completes, provide the
file:// download link
- Hebrew transcription quality is good but may need manual fixes for technical terms
