abe238/claude-video-plus

This project is a skill that lets AI coding assistants such as Claude Code, Codex, Cursor, and similar tools watch a video and answer questions about it. It is a fork of an earlier project called claude-video, meaning it started from that project's code and license and then added new features on top, with credit kept to the original author. The original tool answers a question about a video by sampling frames and reading the transcript across the entire video, no matter what was asked, which can use a large number of tokens, the units AI models are billed and limited by. This fork adds an optional new mode that only turns on when a specific question is given. In that mode, it looks for whole topical chapters, catches numbers like prices or specifications wherever they appear with a matching frame, and grabs frames at chapter starts or moments the video seems to be pointing something out. According to the project's own benchmark, run on five videos the tool had not seen before with answers checked in advance by human reviewers, this question-focused mode gave answers judged just as good as the original approach while sending the AI 56 percent less material and finishing about 14 seconds faster. If there are no captions, the video is a local file, it runs under about 9 minutes, or anything goes wrong, the tool automatically falls back to the original full-video approach, so there is no case where the new mode gives a worse result than before. No question means the original mode runs as before. Setup requires no account or API key. Missing tools like yt-dlp and ffmpeg install themselves automatically the first time it runs. When a video has no captions, the tool tries to transcribe the audio locally on the user's own machine first, and only sends audio to a cloud transcription service if the user explicitly turns that option on. The README also describes specific steps taken to stop text pulled from a video, such as its title or description, from being able to trick the AI into following hidden instructions. The project is released under the MIT license.