--- name: Read Anything slug: read-anything category: Writing description: Analyze any link or local file—web pages, videos, podcasts, social posts, meeting recordings—and get structured Chinese summaries (detailed notes, quick scans, interview breakdowns, meeting minutes, paper digests, content ideas) plus follow-up Q&A based on the source material. github: "https://github.com/heyuxuan0209/read-anything" language: JavaScript stars: 25 forks: 3 install: "npx degit https://github.com/heyuxuan0209/read-anything ~/.claude/skills/read-anything" installs_to: ~/.claude/skills/read-anything source_path: SKILL.md collection_size: 1 category_size: 1012 added: 2026-08-11T07:22:00.390Z last_synced: 2026-08-11T07:22:00.390Z canonical_url: "https://dirskills.com/skills/read-anything" --- # Read Anything Analyze any link or local file—web pages, videos, podcasts, social posts, meeting recordings—and get structured Chinese summaries (detailed notes, quick scans, interview breakdowns, meeting minutes, paper digests, content ideas) plus follow-up Q&A based on the source material. **Install:** ```bash npx degit https://github.com/heyuxuan0209/read-anything ~/.claude/skills/read-anything ``` ## README # read-anything · 读不完的收藏，让 AI 替你读懂 > Too much saved, too little read — let your AI read it for you. **English** · [中文](#中文)
read-anything 演示：网页/长文 → 一键读成中文结构化笔记 演示：丢一个链接进去 → 自动抓取 / 转写 → 出中文结构化解读（精读稿 / 快扫 / 访谈拆解…）→ 就着材料继续追问。
A link in → auto-fetch / transcribe → a structured Chinese explainer → keep asking follow-ups over the source.
``` 你：https://www.xiaoyuzhoufm.com/episode/xxxx 这期讲了啥 它：（本地转写音频 + shownotes → 访谈拆解：两位嘉宾各自观点、分歧点、金句带时间戳） 你：他们在 Agent 记忆那段具体怎么说的？ 它：（只基于转写回答，标时间戳；没提到的如实说没提到） ``` ## English Is your bookmarks folder full of things you "plan to get to someday" — tens of thousands of words of long-form deep dives, an hour-long English podcast, a forty-minute Bilibili video, dozens of X threads… that just gather dust because you can't finish them, or can't even get through them? read-anything helps you understand them fast: drop in a link, and a few minutes later you get a Chinese explainer that spells out "what this actually says," and you can keep asking follow-up questions about it. It all happens inside your AI agent, the results land as markdown, and they accumulate into your own knowledge base. ### The specific hassle it solves - **Saved it but no time to read** → in a few minutes you get an explainer that "makes sense even without the original" - **It's English / video / audio, and a pain to get through** → auto-transcription + a structured Chinese explainer straight out - **Want to dig into details** → converse over the material; answers are based only on the source, never made up - **Read it and forgot it, scattered everywhere** → explainers land as markdown, accumulating into your own system - **Don't want to bounce between a dozen web tools** → one link, all inside your agent ### Formats it can handle Web pages / blogs / single WeChat public-account articles / YouTube / Bilibili / Xiaoyuzhou podcasts / X posts / **local recordings** (meeting recordings, voice memos — just drop the file path for m4a/mp3/wav) / plain text (including pasted meeting notes, speech-to-text transcripts). English content produces a Chinese explainer directly (no need to translate first and then read); videos, podcasts, and local recordings without captions go through local transcription (faster-whisper — audio never leaves your machine, zero API cost). Document links behind a login wall like Feishu/Notion can't be fetched (it will tell you so honestly) — just export the text transcript from Miaoji and paste it in, and it will break it down all the same. ### Six reading perspectives (+ your own) | 你想要 | 用哪个模板 | |---|---| | 完整读懂，不看原文也行 | 精读稿（默认） | | 30 秒判断值不值得细读 | 快扫 | | 播客 / 对谈，按人拆观点 | 访谈拆解 | | 会议录音/纪要，拆决策·待办·分歧 | 会议纪要 | | arXiv / 技术报告 | 论文拆解 | | 从内容里挖可写的选题和素材 | 选题素材 | Not enough? Drop one of your own .md files into `templates/` and it becomes a new template — just say "use my xx template." ### One principle: if it can't get it, it says so Fetch failures, only getting a summary, audio that didn't transcribe — it tells you all of it honestly, and never forces out a polished-looking explainer from incomplete material. Transcription output is labeled "may contain transcription errors," and quoted highlights remind you to check against the original. Every sentence you read is traceable. ### Install **Claude Code** (one command): ```bash git clone https://github.com/heyuxuan0209/read-anything.git ~/.claude/skills/read-anything ``` Once installed, in any session just drop a link and say "analyze this" to trigger it; update with `git -C ~/.claude/skills/read-anything pull`. To install it only for a specific project, clone into that project's `.claude/skills/read-anything`. **Other agents (Codex / Cursor / Gemini CLI / OpenClaw…)**: the core logic all lives in `PLAYBOOK.md` (a pure-markdown playbook) + `scripts/` (plain command-line tools, JSON in and out), with no dependency on any Claude feature. The way to wire it into any agent follows the same pattern — **clone + paste a paragraph into your agent's instruction file**: ```bash git clone https://github.com/heyuxuan0209/read-anything.git ~/.agent-skills/read-anything ``` Paste this into the file below (only the filename differs across agents): ```markdown ## read-anything（链接/录音 → 中文解读） 当我丢链接、音频文件，或粘贴长文/纪要并要求「分析 / 解读 / 讲了啥」时， 按 ~/.agent-skills/read-anything/PLAYBOOK.md 执行（路由、模板、降级规则都在里面）。 ``` | Agent | 指令文件 | |---|---| | Codex CLI | `~/.codex/AGENTS.md`（全局）或项目根 `AGENTS.md` | | Cursor | 项目 `.cursor/rules/` 下新建规则（或旧版 `.cursorrules`） | | Gemini CLI | `~/.gemini/GEMINI.md`（全局）或项目根 `GEMINI.md` | | 其他 / 懒得配 | 每次直接说「按 ~/.agent-skills/read-anything/PLAYBOOK.md 处理这个链接」 | Honest boundary: the scripts guarantee every agent gets the same clean material; **the quality of the explainer is determined by the model you use**, and consistent results across models is not promised. ### Dependencies | 依赖 | 是否必需 | 用途 | |---|---|---| | Node.js 18+ | 必需 | 三个抓取脚本的运行时（零 npm 依赖） | | yt-dlp | 可选 | YouTube/B站（`brew install yt-dlp`） | | faster-whisper | 可选 | 播客/无字幕视频本地转写（`pip install faster-whisper`，首次自动下载 ~460MB 模型） | Missing an optional dependency won't cause an error — the corresponding content degrades automatically and **explicitly tells you it degraded**. Accessing YouTube from within China requires setting the `YOUTUBE_PROXY_URL` environment variable (e.g. `http://127.0.0.1:7890`). ### File structure ``` read-anything/ ├── SKILL.md # Claude Code 触发入口（薄壳） ├── PLAYBOOK.md # 通用剧本：路由/材料规范/降级链/诚实守则（agent 无关）★ ├── templates/ # 解读模板（精读/快扫/访谈/会议/论文/选题素材 + 你自己的） └── scripts/ # fetch-article / fetch-xiaoyuzhou / vtt-to-text（零依赖）+ transcribe.py（可选） ``` ### Relationship to NotebookLM It can converse around your material the way NotebookLM does, but without uploading to any cloud service, without switching web pages; audio is transcribed locally and never leaves your machine, and the output is your own markdown file. ### Origins Distilled from the "instant analysis" pipeline of the personal knowledge management system [knowledge-workbench](https://github.com/heyuxuan0209/knowledge-workbench) — that one is a full "collect → understand → accumulate → create" workbench, and this is the standalone, shareable version of its "understand one piece of content" step. The hardening points in the scripts (seeing through Jina's fake 200s, the 300-character junk threshold, mis-heard-name warnings for captions, the podcast-without-transcript disclaimer, etc.) all come from real pitfalls, with the reasons kept in the comments. --- ## 中文 收藏夹里是不是躺着一堆「打算有空再看」的东西——几万字的深度长文、 一小时的英文播客、四十分钟的 B 站视频、几十条的 X thread…… 结果一直吃灰，读不完也读不动。 read-anything 帮你把它们快速读懂：丢一个链接进来，几分钟后拿到一篇 讲清楚「这到底说了什么」的中文解读，还能就着它继续追问。全程在你的 AI agent 里完成，解读结果落成 markdown，沉淀进你自己的知识库。 ### 它解决的那个具体麻烦 - **收藏了没时间读** → 几分钟给你一篇「不看原文也懂」的解读 - **是英文 / 视频 / 音频，读起来费劲** → 自动转写 + 直接产出中文结构化解读 - **想追问细节** → 就着材料对话，答案只基于原文，不编造 - **读完就忘、散落各处** → 解读落成 markdown，沉淀进你自己的系统 - **不想在十几个网页工具间来回切** → 一个链接，全程在你的 agent 里 ### 吃得下这些格式 网页 / 博客 / 公众号单篇 / YouTube / B站 / 小宇宙播客 / X 帖子 / **本地录音**（会议录音、语音备忘录，m4a/mp3/wav 直接丢文件路径）/ 纯文本 （含粘贴的会议纪要、语音转文字稿）。 英文内容直接产出中文解读（不用先翻译再读）；无字幕的视频、播客和本地 录音走本地转写（faster-whisper，音频不出本机、零 API 费）。 飞书/Notion 等登录墙后的文档链接抓不到（会如实告诉你）——从妙记导出 文字记录粘贴进来即可，照样拆解。 ### 六种解读视角（+ 你自己的） | 你想要 | 用哪个模板 | |---|---| | 完整读懂，不看原文也行 | 精读稿（默认） | | 30 秒判断值不值得细读 | 快扫 | | 播客 / 对谈，按人拆观点 | 访谈拆解 | | 会议录音/纪要，拆决策·待办·分歧 | 会议纪要 | | arXiv / 技术报告 | 论文拆解 | | 从内容里挖可写的选题和素材 | 选题素材 | 不够用？往 `templates/` 丢一个你自己的 .md 就是新模板，说一句「用我的 xx 模板」即可。 ### 一个原则：拿不到就说拿不到 抓取失败、只拿到摘要、音频没转成——都如实告诉你，绝不拿残缺材料硬编 一篇像模像样的解读。转写产物标注「可能存在听写误差」，引用金句提醒你 核对原文。你读到的每一句都有据可查。 ### 安装 **Claude Code**（一条命令）： ```bash git clone https://github.com/heyuxuan0209/read-anything.git ~/.claude/skills/read-anything ``` 装完在任何会话里丢个链接说「分析一下」即可触发；更新用 `git -C ~/.claude/skills/read-anything pull`。只装到某个项目就克隆到 项目的 `.claude/skills/read-anything`。 **其他 agent（Codex / Cursor / Gemini CLI / OpenClaw…）**：核心逻辑全在 `PLAYBOOK.md`（纯 markdown 剧本）+ `scripts/`（普通命令行工具，JSON 进出）， 不依赖任何 Claude 特性。所有 agent 的接法都是同一个模式——**克隆 + 在你的 agent 指令文件里贴一段话**： ```bash git clone https://github.com/heyuxuan0209/read-anything.git ~/.agent-skills/read-anything ``` 往下面这个文件里贴这段（各家只是文件名不同）： ```markdown ## read-anything（链接/录音 → 中文解读） 当我丢链接、音频文件，或粘贴长文/纪要并要求「分析 / 解读 / 讲了啥」时， 按 ~/.agent-skills/read-anything/PLAYBOOK.md 执行（路由、模板、降级规则都在里面）。 ``` | Agent | 指令文件 | |---|---| | Codex CLI | `~/.codex/AGENTS.md`（全局）或项目根 `AGENTS.md` | | Cursor | 项目 `.cursor/rules/` 下新建规则（或旧版 `.cursorrules`） | | Gemini CLI | `~/.gemini/GEMINI.md`（全局）或项目根 `GEMINI.md` | | 其他 / 懒得配 | 每次直接说「按 ~/.agent-skills/read-anything/PLAYBOOK.md 处理这个链接」 | 诚实边界：脚本保证各家 agent 拿到同样干净的材料；**解读质量由你所用模型决定**， 不承诺跨模型效果一致。 ### 依赖 | 依赖 | 是否必需 | 用途 | |---|---|---| | Node.js 18+ | 必需 | 三个抓取脚本的运行时（零 npm 依赖） | | yt-dlp | 可选 | YouTube/B站（`brew install yt-dlp`） | | faster-whisper | 可选 | 播客/无字幕视频本地转写（`pip install faster-whisper`，首次自动下载 ~460MB 模型） | 缺可选依赖不报错——对应内容自动降级并**显式告知你降级了**。 国内网络访问 YouTube 需设置 `YOUTUBE_PROXY_URL` 环境变量（如 `http://127.0.0.1:7890`）。 ### 文件结构 ``` read-anything/ ├── SKILL.md # Claude Code 触发入口（薄壳） ├── PLAYBOOK.md # 通用剧本：路由/材料规范/降级链/诚实守则（agent 无关）★ ├── templates/ # 解读模板（精读/快扫/访谈/会议/论文/选题素材 + 你自己的） └── scripts/ # fetch-article / fetch-xiaoyuzhou / vtt-to-text（零依赖）+ transcribe.py（可选） ``` ### 和 NotebookLM 的关系 能像 NotebookLM 那样围着材料对话，但不用上传给任何云服务、不切网页， 音频本地转写不出本机，产物是你自己的 markdown 文件。 ### 出处 从个人知识管理系统 [knowledge-workbench](https://github.com/heyuxuan0209/knowledge-workbench) 的「即时分析」链路提炼——那边是完整的「采集 → 理解 → 沉淀 → 创作」工作台， 这里是其中「读懂一条内容」环节的独立可分享版。脚本里的加固点（Jina 伪 200 识破、300 字垃圾门槛、字幕误听人名警示、播客无转写声明等）都来自真实踩坑， 注释里保留了原因。 ## 🔗 关注我 · Follow me 边做 AI 产品边把一手经验和思考公开分享，欢迎关注、来聊。
I build AI products in public and share the notes here — come say hi:
小红书 · Xiaohongshu 公众号 · WeChat 视频号 · Channels 抖音 · Douyin
小红书 杰西卡 公众号 杰西卡聊AI 视频号 杰西卡 抖音 杰西卡
## License & 二开须知 · Contributing MIT — 见 [LICENSE](LICENSE)。欢迎 **Star / Fork / Issue**，也欢迎二次开发、魔改、接进你自己的产品或工作流。**唯一的请求**：二开或转载时**注明出处**，并 **@ 一下我**（公众号 / 小红书「**杰西卡聊AI**」，主页见上）——让顺着来的人能找到源头，就是最好的感谢 🙏。 MIT licensed — see [LICENSE](LICENSE). **Star / Fork / Issues welcome**, and feel free to remix, modify, or build it into your own product or workflow. **One ask:** if you fork/remix or repost, please **credit the source and @ me** (Jessica · 杰西卡聊AI). That's the best thank-you 🙏.