joeseesun/qiaomu-cut-skill

qiaomu-cut is a video editing skill for AI agents. You describe the video you want in one sentence, and it organizes the entire production process: sourcing footage, planning shots, adding subtitles, transitions, motion graphics, rendering, and quality checks. The goal is to turn what is normally a messy, manual workflow into a structured, reproducible system that a software agent can execute and verify. The project works by converting your request into an intermediate representation called QiaoCut IR, which captures details like duration, aspect ratio, audience, style, shot list, and rendering strategy. From there, it routes work to various tools and sources: stock footage providers, local files, text-to-speech and music services, speech recognition, AI image generation, and rendering engines like ffmpeg. Version 0.5 includes project timeline rendering, bilingual subtitles with large mobile-friendly text, brand intro and outro templates, file and scene deduplication checks, and three rendering quality levels called preview, standard, and final. The README is honest about what is actually built versus what remains a work in progress. Features like HTML video capture, Manim and PowerPoint export, complex masking, speed ramps, and a full transition library are described as future directions rather than finished capabilities. The project also takes care with paid operations: tasks that cost credits or require uploads always need explicit confirmation, and it never silently approves charges on your behalf. Installation is a single command using npx, and a doctor command checks whether your machine has the required tools like Node.js and ffmpeg. The documentation includes many command-line examples for searching footage, generating narration, transcribing audio, planning a video, scaffolding a project, and rendering at different quality tiers. Security practices are notably careful: credential files must have restrictive permissions, temporary URLs are kept out of project manifests, and downloads are validated for content type and file integrity. The full README is longer than what was shown.