Claude Shot Catalogue — Development skill for Claude Code

shot-catalogue

A [Claude Agent Skill](https://docs.claude.com/en/docs/agents-and-tools/agent-skills/overview) that turns a folder of rushes into a document an editor can cut from.

The value isn't the file list — the operating system already has that. It's the four questions a file list can't answer:

• Which sound goes with this picture, and at what offset?
• Which script beats does this take actually cover?
• What was changed or dropped on the day?
• What did we never get at all?

Everything runs locally. Transcription uses Whisper in a throwaway virtualenv; sync uses ffmpeg and numpy. No API keys, no per-file cost, no upload of your footage anywhere.

Install

Clone into your skills directory:

git clone https://github.com/VehpuS/claude-shot-catalogue.git ~/.claude/skills/shot-catalogue

Then ask Claude to catalogue a footage folder. The skill triggers on phrasings like *"log these rushes"*, *"which take covers this line"*, *"which WAV goes with which clip"*, *"what coverage are we missing"* — it doesn't need the word "catalogue".

**Requirements:** `ffmpeg`/`ffprobe`, Python 3 with `numpy`. Whisper installs itself on first run (`mlx-whisper` on Apple Silicon, `faster-whisper` elsewhere). `pdftotext` only if your script is a PDF.

How the sync works

Pairing dual-system sound to picture is where naive approaches produce confident, wrong answers. Two signals are needed, because each fails exactly where the other holds:

	Tells you	Fails when
Transcript overlap	Identity — which recording holds this performance	Takes repeat the same lines, which is what takes do
Loudness-envelope correlation	Timing — the exact offset	Searched unconstrained: latches onto spurious peaks

So the matcher shortlists candidates by text, derives a coarse offset from the median time difference across matched utterances, then refines with a normalized cross-correlation **constrained to a few seconds around that anc