Project
Case-to-video pipeline
An internal tool that turns a court or personal rehabilitation case report into a short-form or long-form YouTube video draft. Claude Opus writes the script, and every stage waits for a person to confirm it.
How it works
- Read the report. We upload a case report PDF (HWP and HWPX files also work, and an existing YouTube video can be split into several Shorts). The tool extracts the case facts, and a person confirms them.
- Write the script. Claude Opus plans and writes the script, 12 to 16 content turns for a Shorts episode. The tool calls Claude through the Claude Code CLI. Personal details in the report are masked before any text is sent.
- Record the voices. Two-voice narration, a host and a lawyer character, is generated with a text-to-speech engine.
- Draw the visuals. Four to five handwritten-style image cards per Short, plus detail cards that show key facts quoted from the report.
- Caption and render. Speech is aligned with Whisper so captions match the voice, and the video is rendered in 9:16 or 16:9.
Each stage needs a person to press confirm before the next one runs, and the finished video is reviewed before it is published.
Formats
- Conversation Shorts, about 75 to 80 seconds
- Single-voice meme Shorts
- Splitting one long video into several Shorts
- Live-action Shorts
- Search-intent Shorts written around a viewer's question
- Long-form 16:9 explainers
Privacy by design
Case reports are masked before any text reaches a language model. In the long-form writer mode only an anonymized summary is sent, and amounts, periods and dates come back as placeholders that are filled in locally. If a model's output contains sensitive data or numbers that are not in the report, the tool discards it and falls back to a template.
Engineering
- Python, ffmpeg and faster-whisper, run as a local web app
- 1,465 regression tests, including a check that tests never touch paid voice storage or production logs
- In development since July 16, 2026, with 271 commits as of October 2026