threerocks/article2media
Productivity & UtilitiesTurns an article or video into publish-ready Xiaohongshu cards, long images, WeChat layouts and narrated shorts via Claude Code or Codex.
What it does
article2media is a set of Claude Code / Codex skills, also shipped as a Claude Code plugin, that takes one article or one video and produces Chinese visual content ready to publish: phone-width long images, Xiaohongshu (RedNote) cards, WeChat Official Account layouts, 9:16 narrated videos, plus verbatim re-layouts, full translation, an argument whiteboard and a comic. It is built as ten independent skills that hand off files in a shared working directory, so you can run only the step you need or chain them.
The division of labor matters: the model handles understanding and writing (rewrite, translation, video script, whiteboard ledger, storyboard), while scripts do the deterministic work: fetching, HTML layout, headless Chrome screenshots, voice-over, word alignment, mixing and validation. Generated images never contain text; all text is laid out in HTML, which keeps output checkable and editable.
Quick start
Prerequisites: Node.js 22 and a Chrome/Chromium installation. Install the plugin in Claude Code, which brings in all skills at once:
claude plugin marketplace add threerocks/article2media
claude plugin install article2media@article2media
In the repository, run the environment check and optionally download theme fonts locally (about 65 MB):
bash scripts/install.sh
node skills/article2media/scripts/fonts.mjs
Then invoke the entry skill with a URL or local file and the desired output:
/article2media https://example.com/post 小红书
/article2media ./my-essay.md 公众号
/article2media https://www.bilibili.com/video/BV… 小红书
Outputs land in article2media-out/<slug>/ under the current directory. Individual skills can be called directly, for example /article2media:wechat-layout ./my-essay.md for layout only, without summarization. Codex users invoke $article2media instead. Command-line script entry points are listed in the README for running without a model host.
Outputs and capabilities
The default request produces a summarized long image. Cards default to a 3:4 Xiaohongshu set with six cover layouts; --ratio switches to 4:5, 1:1 or 9:16. WeChat layout takes the original, translated or exported rewritten article and produces inline-styled HTML that pastes into the editor, plus header and cover images. Verbatim paths (-f, 原文卡片) keep every block unchanged and read it back for comparison. Translation outputs article.zh.md with a block-structure check. The explainer video renders a 1080×1920 narrated video with local VoxCPM2 voice (or edge-tts / system voice fallback), word-synced captions and a review report. Video input is first transcribed: platform subtitles when available, otherwise local Whisper on Apple Silicon.
There are 12 main themes and 9 legacy themes; fonts load from Google Fonts and jsDelivr by default unless mirrored locally, and the layout checker will name any font family that failed to load.
Constraints
- Node.js 22+ and Chrome/Chromium are required for all outputs. ffmpeg is required for narrated video.
- Local voice quality depends on Apple Silicon and VoxCPM2; other environments fall back to edge-tts (network) or the macOS system voice.
- AI illustrations and comics need an image provider: a logged-in Codex CLI, or OpenAI / Gemini paid keys. Video transcription and word alignment without subtitles need mlx-whisper. YouTube/Bilibili subtitles and Douyin downloads need yt-dlp.
- Some sites may require browser cookies; the scripts offer
--cookies-from-browser chromeor a local file instead. - Publish statistics are stored only in the local working directory; comparisons with fewer than 5 samples are marked as insufficient.
- Cloned voices stay in the local voice library and are not distributed; videos carry AI-generated labels. Code is MIT licensed; bundled example content belongs to its original authors.
HotRepo's take
On the 2026-11-11 new-project ranking for JavaScript, the repository recorded 24 stars in roughly 90 minutes of measured growth, reaching 126 stars and 25 forks overall. That is a single snapshot, so it supports only the fact of a fast start, not a trend. What stands out from the materials is the engineering discipline behind the outputs: verbatim block ledgers with SHA-256 source locking, per-output acceptance checks that exit non-zero, and a test matrix covering examples across themes and layouts. The project is also unusually explicit about what is optional and what degrades gracefully, which matters for anyone evaluating it against a paid design tool.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- andyhuo520/vibe-knowledgeA Claude Code skill that turns one knowledge point into a narrated, sourced short explainer video, rendered frame by frame from code.
- Vincentwei1021/anything2explainerTurns any topic into a narrated, code-drawn explainer video through a Claude Code or Codex skill.
- mu-zi-lee/TangentryAggregates videos, podcasts, RSS and articles into one local reader where notes anchor to timestamps and text.