A developer used a five-year-old MacBook Pro M1 Max with 64GB RAM to index a year of unlabeled video footage locally, running Gemma 4 31B through LM Studio and peaking at 50.89 GB of swap during the bulk pass. The build leveraged Claude Code to orchestrate metadata extraction, frame sampling, transcription, and face detection across clips stored on physical SSDs.

The pipeline extracts five frames per clip with ffmpeg, runs WhisperX for word-aligned transcripts across 97 languages, and uses insightface to store 512-dim ArcFace embeddings in SQLite. Each clip gets a sidecar markdown file with YAML frontmatter and prose description, making the archive queryable without uploading to a third party. The whole skill runs from a ~1,400-line Python directory in Claude Code.

Most AI video editors assume labeled footage. This workflow proves the real bottleneck is the index, not the editor. Local-first architecture with sidecars means the archive survives if the indexer breaks, and costs dropped from $140 monthly to $22 by ditching SaaS tools.