Turns screen recordings into editable SOPs. Auto-updates docs from new product videos:
- Upload video
- AI extracts meaningful frames
- Find screenshot worthy timestamps with transcription + AI
- Find out spike in frame changes and note the timestamps
- Merge results from (a) and (b) to obtain screenshot worthy frames
- Find stable frames for the timestamps obtained
- Claude structures it into steps with screenshots
- Edit in a Notion-style TipTap editor
- Regenerate from a new updated recording
- Get a visual diff of what changed.
I Built this to explore the same problem space Clueso works on, end to end
- Fire-and-forget background jobs - Volatile
- Loading-screen/transitional frames can still slip into both frame selection - mitigated in layers, not eliminated
- Requires the user to re-record the entire video for feature update
- Version 0.0.2 → Incremental updates via short clips, not full re-recordings: instead of re-uploading an entire workflow, let the user record just the part that changed (a new step, a modified screen) and splice it into the existing document at the right position.
- Version 0.0.3 → GitHub Webhook to auto-update the documentation: Pushes to github triggers a workflow to look for new changes in the product, record the new feature and auto update the docs.
Presigned S3 uploads (bypass the API tier for large files), SQS-backed job queue with DLQ/redrive for durability and retry, Fargate worker fleet auto-scaled on queue depth (chosen over Lambda for ffmpeg's disk/runtime needs), Postgres + pgvector for step embeddings at real scale, Redis for OTP rate-limiting and diff-result caching, CloudFront in front of private S3 for delivery.