Design#Text to speech#Video generation#Motion graphics
vibe-knowledge: a Claude Code skill for sourced explainers
vibe-knowledge is a Claude Code skill for sourced explainers: fact checks, narration and 1080p60 rendering, with optional Renoise credits for selected shots.
Skill details
npx skills add andyhuo520/vibe-knowledge --skill vibe-knowledgeA knowledge video usually breaks down between research, narration, visuals and editing. vibe-knowledge is andyhuo520’s Claude Code skill for carrying one topic from fact checks through a narrated, frame-rendered explainer. Its rules put every name, number and quotation against a source, keep the narration responsible for the explanation, and reserve on-screen text for keywords, figures and citations. The repository lists 44 visual styles across eight families and targets 16:9, 1080p60 output.
How the Claude Code skill builds a video
This is a guided production workflow, not a one-line prompt that hands back a finished film. The user chooses a topic, one of six storytelling templates, a duration, a visual style and a production tier. A small scoring script weighs concept count, background knowledge, causal steps, evidence density and disagreement, then suggests one of four runtimes: quick, focused, standard or deep. The templates cover a puzzle-led concept, a historical or psychological dilemma, an everyday problem explained by a principle, a through-line across a history, a challenge format and a six-sided rotation of viewpoints.
The skill then creates a facts.md ledger and asks for sources behind every factual claim; competing historical accounts stay visible as competing accounts. A narration draft follows the chosen template and gets an adversarial review for its hook, evidence and length. A recommender reads the repository’s catalog.json and proposes three visually distinct styles for the subject. Before any generated assets are made, the user reviews the narration, storyboard and estimate. A final review happens after rendering, before the package of title, caption and cover is assembled.
Code first, generation for key shots
The base tier uses no Renoise credits. Claude Opus writes layered SVG artwork, while the cine engine supplies particles, charts, map effects and 3D scenes. The know.js kit keeps these elements aligned to narration and includes counters, source footnotes, quote cards, trails and choice buttons. The production rules call for movement at least every three seconds, a consistent visual system and an OCR pass that checks visible wording against the facts ledger. AI-generated people and scenes need an “AI illustration” label.
The enhanced tier can send a few photorealistic scenes or memorable moving shots to Renoise; the remaining work stays code-rendered, so one film.js timeline supports both tiers. The skill requires an itemized estimate and the user’s explicit approval before each generation batch. Changing the model, prompt or quantity means getting a new estimate and approval. Treat the estimate returned at run time as current; credit amounts in repository examples are snapshots. The repository’s own background track is the default for every video unless the user asks to replace it.
Requirements and rights
A complete setup needs Claude Code, the companion ai-teardown skill, a Python environment with numpy, scipy, Pillow and Playwright, plus headless Chromium. Rendering also uses ffmpeg; OCR relies on macOS swiftc and Apple Vision. Narration goes through Xiaomi MiMo, using the key in ~/.config/ai-teardown/mimo.env. SKILL.md labels the MiMo voice service as free for a limited time; check the provider’s current terms. Enhanced shots additionally need a Renoise MCP account and credits; the base tier does not need Renoise. Following SKILL.md, new_film.py creates an episode’s files in the workspace, so it is not a read-only preview. This entry is based on the repository’s README, SKILL.md and NOTICE; the full pipeline was not run end to end here.
As of November 12, 2026, GitHub showed 634 stars and identified the code license as MIT. The project’s NOTICE narrows that license: it covers code and documentation, while sample videos, posters, the style wall, animations and each style’s still.jpg are the author’s work and must not be re-uploaded. The bundled BGM may be used in videos made with the tool but cannot be redistributed or sold by itself. That rules out the repository’s sample art as a Frontier Now cover.
I’d choose this when I want one Claude Code workflow to carry a researched topic through narration, visuals, quality checks and delivery, and I’m willing to keep the local rendering stack working. For a math-first animation, manim-video builds the explanation around Manim scenes. For a narrower narrated drawing workflow, sweety-handdraw-video draws in time with the voice. vibe-knowledge covers more of production, but it asks for more setup and leaves fact checking and media permissions with the creator.