Hockey Film Review

Drop in short Hudl clips of one player, get a coaching report as a single HTML file (annotated stills inlined, responsive) that can be emailed and read on an iPad.

Pipeline: ffmpeg samples frames per clip → Claude vision analyzes each clip (good / fix moments with timestamps, themes, coaching note, key moments to annotate) → stills are annotated with Pillow → a second Claude pass writes the cross-clip summary (strengths, areas to improve, why it happens, development plan) → report.html.

Setup

Requires Python 3.11+, ffmpeg/ffprobe on PATH, and an Anthropic API key.

pip install -r requirements.txt
export ANTHROPIC_API_KEY=sk-ant-...

Use

1. Put the clips in a folder (they are numbered in filename order). Optional clips.csv in that folder: filename,title,url — title is the coach's Hudl tag, url is a per-clip Google Photos link.
2. Copy examples/game.toml, set the player, jersey color, game title and album link.
3. Run:

python -m film_review ./clips --config game.toml --out ./out

Result: out/report.html. Attach it to an email. Flags:

• --album URL Google Photos album link (shown at the top of the report)
• --embed-videos embed 360p copies of each clip so they play inside the file (much larger file)
• --model ID override the model
• --mock offline test with fake analysis, no API call

Per-clip analyses are cached in out/.work, so re-runs only pay for new or changed clips.

Notes

• Claude sees sampled stills (default 3 fps), not continuous video. The report flags clips where the jersey number could not be read and the player was placed by position.