vibe-knowledge skill (Claude Code) + 小米 MiMo 配音 + Renoise(可选)Partly tested🇨🇳 CN

The Douyin "Vibe Knowledge" explainer-short format as an open-source Claude Code skill: give it a topic and it fact-checks, writes, voices, renders frame by frame in code and runs QA; the base tier costs zero credits

Chinese creator @berryxia turned the viral Douyin #Vibe知识大赏 ("Vibe Knowledge") explainer-short format into an open-source skill, vibe-knowledge: 6 storytelling templates (guess-the-concept, dilemma, everyday-paradox, etc.), 44 visual styles, and every narration line must trace to a source in facts.md. The base tier is pure Claude Opus SVG/code animation at zero credits; the enhanced tier hands realistic shots to Renoise (quote first, then generate). The repo ships 7 sample films. We cloned the 2026-10-28 evening build on Linux: the style recommender works, but workspace setup stopped at the BGM patch added that same day; we rendered and voiced nothing.

Visit site vibe-knowledge skill (Claude Code) + 小米 MiMo 配音 + Renoise(可选) ↗
Pricing
Freemium
Region
Global
Where it's from
Chinese community
Added
Oct 29, 2026
Tested on
Oct 29, 2026
Version we used
andyhuo520/vibe-knowledge commit c3a13a8 (2026-10-28 23:30 +08); CHANGELOG 2026-10-28 afternoon (BGM forcing); 44 styles
Categories
VideoWritingResearch
Original author
Berryxia.AI @berryxia ↗
Our test
Partly tested
Cost / time
The skill is MIT-licensed and free; you need your own Claude Code (or other agent) plan; Xiaomi MiMo voice is free for now per the author; the enhanced tier spends Renoise credits, the base tier spends none.
Source
X @berryxia(开源发布) ↗X @berryxia(预告:43 套风格、6 种讲法、两档制作) ↗GitHub andyhuo520/vibe-knowledge ↗

The task

Make polished "guess first, then explain" science/history shorts for Douyin or Bilibili knowledge channels without editing or animation skills—and without the AI making up facts.

Steps

  1. Install: `git clone https://github.com/andyhuo520/vibe-knowledge.git`, then `cp -R vibe-knowledge/skills/vibe-knowledge vibe-knowledge/skills/ai-teardown ~/.claude/skills/`. You need both: ai-teardown supplies the cine render engine, workspace setup and voice scripts.
  2. Render environment: `python3 -m venv ~/.venvs/cine && ~/.venvs/cine/bin/pip install numpy scipy pillow playwright && ~/.venvs/cine/bin/python -m playwright install chromium-headless-shell`, then `export CINE_PY=~/.venvs/cine/bin/python`; plus ffmpeg. OCR QA uses macOS's built-in Vision (needs swiftc).
  3. Voice: Xiaomi MiMo; put the key in `MIMO_API_KEY` or `~/.config/ai-teardown/mimo.env` (never in the project). The default voice is an AI voice bundled with the repo; to use your own, run `new_film.py --voice-ref your-recording.wav` (only your own or an authorized voice).
  4. Enhanced tier (optional): connect Renoise MCP—for Codex `codex mcp add renoise --url https://renoise.ai/api/public/mcp` then `codex mcp login renoise`; for Claude Code `/plugin marketplace add ArcoCodes/renoise-plugins-official`. The base tier needs none of this.
  5. In Claude Code say "做一期知识大赏:<your topic>" ("make a Vibe Knowledge episode: …"). First run opens a local style gallery; `tools/recommend.py` suggests 3 visually distinct styles from topic keywords.
  6. Checkpoint 1: topic, template (1 of 6), length tier (quick / focused / standard / deep), style, production tier.
  7. Research and script: it writes facts.md first (every number, name, year and quote sourced; disputed claims shown side by side), then narration per template, with a sub-agent adversarially reviewing the hook, sources and length. Checkpoint 2: approve script and storyboard (enhanced tier adds a shot list and Renoise quote) before anything is generated.
  8. Voice → film.js composition → 1080p60 frame-by-frame capture → mix (mandatory default BGM, 10 dB under the voice) → QA (contact sheet, OCR read-back against facts.md, loudness, freeze and BGM detection). Checkpoint 3: review the cut and get a publish pack (title, hashtags, description, cover).

Prompt

Make a Vibe知识大赏 (Vibe Knowledge) episode: how many times would you have to fold a sheet of paper to reach the Moon? Recommend 3 styles and a length tier first, use the zero-credit base tier, source every number in facts.md, write the narration in English, and show me the script and storyboard for approval before voicing and rendering.

What the author says

@berryxia: name a topic and you get a polished explainer short with narration where every line is sourced; 43 styles in 8 families (44 in the repo now) and 6 storytelling templates; pure code renders at zero credits, add AI images only for realistic shots; works with Claude, Codex, WorkBuddy, Doubao, Qwen Work and other agents. The README says episode 01 (Shang Yang's log) cost 246 Renoise credits in the enhanced tier (10 Midjourney images + 2 Wan motion shots + music + 5 SFX), with enhanced runs typically ≈25–40 / 250–300 / 600–700 credits for light / standard / flagship; the release has 7 sample films, including a base-vs-enhanced side-by-side of the same episode.

Our test

On Linux (isolated HOME) we cloned commit c3a13a8 (2026-10-28 23:30 Beijing time), copied both skills into ~/.claude/skills per the README, and ran recommend.py and new_film.py. We did not install Playwright/Chromium, had no MiMo key and did not connect Renoise—so no voice, no rendered video, and no end-to-end episode driven by Claude Code.

What we found when we tried it:catalog.json really lists 44 styles. recommend.py works offline: "商鞅立木为什么能让人相信新法" → gufeng first (matched 商鞅, paired with the dilemma template); "为什么假账总被数字 1 出卖" → no match, fallback to starfield / variety / caption-news. new_film.py 01 --template A built the workspace and series skeleton and patched record.py, stills.py and finish.py, then bgm_force.py --patch exited with an error and lessons/ stayed empty—the BGM-forcing hook added that afternoon doesn't match the finish.py in the bundled ai-teardown.

How to get in

The original post is in Chinese on X; the GitHub repo is public with both a Chinese README and README.en.md. Sample films download directly from the GitHub release. Xiaomi MiMo (voice) and Renoise each need their own account; narration templates are Chinese-first, so ask for English narration explicitly.

Who it suits

Solo creators running knowledge channels on Douyin, Bilibili or YouTube Shorts who want frequent explainer shorts; Mac users on Claude Code who are fine with a few approval rounds.

Pitfalls

  • In our test, the 2026-10-28 evening build (commit c3a13a8) on a fresh install created the workspace but stopped at the BGM-forcing patch added that day (bgm_force.py: "can't find `import hashlib, os` in finish.py… hook needs rewriting"), so lesson 01's narration, film.js and facts.md were never created. The author is iterating fast—check the CHANGELOG and issues before installing.
  • In our test, style recommendation is keyword matching: the README's own example topic ("why fake books get betrayed by the digit 1") matched no style keywords and fell back to the default order (starfield / variety / caption-news); a topic like Shang Yang's log does correctly get the gufeng style.
  • OCR QA relies on macOS Vision (swiftc); on Linux/Windows you need a substitute. Rendering also needs Playwright + Chromium.
  • Voice defaults to Xiaomi MiMo and needs your own API key; only clone your own or an authorized voice.
  • The BGM is mandatory: unless you explicitly ask to change it (quoted in the music entry), other full-length music gets dropped.
  • Renoise links in the repo and post carry referral/UTM parameters, and several creators posted Renoise promos the same week; the enhanced-tier credit numbers are the author's—we generated nothing.
  • "Every line sourced" relies on facts.md and a reviewer sub-agent; you still need to open and check the sources yourself.

More in this category