andyhuo520/vibe-knowledge

This repository packages a Claude Code skill for producing short knowledge explainers: you give it a topic, and the conversation walks through topic selection, sourcing, scriptwriting, voice-over, code-driven rendering, quality checks and a publish package. It suits people already working in Claude Code who want a repeatable pipeline for 1080p60 explainer clips rather than a general-purpose video editor.

How a video is made

There are three explicit confirmation gates: first on topic, master template, duration, style and production tier; then on the narration and storyboard; finally on the finished cut and publish package. Facts are collected into a facts.md file where each number, name, year and quotation carries a source, and disagreements are shown side by side. A sub-agent then reviews the script for hook, sourcing and length before voice-over.

Pictures come from code: a film.js file plus a component library (know.js) covering image pans, keyword callouts, counters, source footnotes, dilemma buttons and diagrams. Rendering is a frame-by-frame 1080p60 capture, followed by mixing with forced BGM calibrated to sit 10 dB below the voice, then checks including contact sheets, OCR read-back against facts.md, loudness, freeze detection and BGM detection.

Two production tiers

The basic tier costs nothing and draws all visuals in code. The enhanced tier sends only the shots code handles badly (realistic period scenes, one to three key dynamic shots, some sound effects) to Renoise, an AI video/image/audio platform connected via MCP. Renoise gives a cost estimate first and only generates after you approve; the README reports the episode 1 enhanced version used 246 credits and gives rough per-tier ranges. Enhanced visuals are labelled "AI 示意" in the film. The same film.js works in both tiers.

Requirements and limits

• Claude Code plus local tooling. Installation copies two skills (vibe-knowledge and ai-teardown) into ~/.claude/skills/. The latter supplies the cine three.js renderer, workspace and voice scripts README installation section.
• Python environment with numpy, scipy, pillow, playwright and headless Chromium, set via CINE_PY; ffmpeg is also needed, and OCR quality checks use macOS Vision via swiftc.
• Voice-over requires a Xiaomi MiMo key in MIMO_API_KEY or a config file, kept out of project files. The default voice is a bundled AI voice; using your own recording requires your own or explicitly authorized voice.
• Renoise is optional and only for the enhanced tier; the basic tier needs no account or credits.
• The README describes the workflow and sample episodes, but does not provide independent quality or accuracy validation; sourcing discipline is a process the skill enforces, not a guarantee of correctness.

Licence

Code is MIT. Sample videos, AI-generated images and shots, the default BGM and default voice have separate terms in NOTICE.md; third-party components and referenced projects are listed in their own files.