GitHub Pick

Clone it, open any AI agent inside the folder, and just say "curate the top GitHub projects on X." You get a localized, double-click-to-open research board — the agent searches GitHub broadly, judges each project's real-world reception, tags it, and writes a localized digest. Zero build, zero API key, nothing to install. The repo is the workspace.

🌏 中文版 · 🖥 Live page · ⚖️ License

中文简介： Clone 下来，在项目文件夹里打开任意一个 AI，直接说一句「帮我聚合 X 主题的开源项目」，就能得到一个本地化、双击即开的开源研究看板。AI 广搜 GitHub、判断每个项目的真实反响、打标签、写本地化点评，一个零构建的静态看板负责展示。完整中文文档 → README_zh.md

---

GitHub Pick turns "I want to learn the open-source landscape of X" into a clean, scannable dashboard. You give a topic; the AI searches GitHub broadly, judges each project's real-world reception, tags it, and writes a localized digest. A zero-build static dashboard (index.html + data.js) renders it — double-click, no server, no build step.

It is not a plugin and not a skill you install anywhere. It is just a project folder with an instruction file (AGENTS.md / CLAUDE.md) that any modern AI coding agent reads automatically. The repo is the workspace.

---

How it works in one picture

clone repo  ──▶  open an AI agent in the folder  ──▶  "帮我聚合 X 主题的开源项目"
                                                              │
        AI reads AGENTS.md + workflow/github-pick.md ◀────────┘
                                                              │
        searches GitHub · tags · writes sessions/*.json ──────┤
                                                              │
        runs `node scripts/build-data.mjs` → data.js ────────▶  open index.html ✦

No installation. No API key. No global config. Everything lives in the cloned folder.

---

Requirements

• An AI coding agent that reads AGENTS.md or CLAUDE.md — e.g. Claude Code, Codex CLI, Cursor, Cline, Aider.
• Node.js (any recent version) — for the tiny zero-dependency build script.
• gh CLI (optional) — if installed and authenticated, GitHub search gets 5000 req/hr instead of the anonymous 60/hr. Either works for a single run.

Quick start

git clone https://github.com/LeifDiao/github-pick.git
cd github-pick
# open your AI agent IN this folder, then just talk:
#   "帮我聚合 RAG 框架 的开源项目"
#   "curate the top Claude Code skill projects"

The first time you run it, the agent asks which language you want. After that the choice is locked (see below) and it never asks again. When it finishes, open index.html.

Prefer to set things up by hand first? Just edit config.json ("lang": "en") and run node scripts/build-data.mjs.

Language — pick once, stay forever

One language per project, chosen on first run, stored in config.json. From then on everything — every digest and the dashboard UI — stays in that language. There is no runtime toggle.

• zh and en ship with built-in dashboard UI labels.
• Any other language (ja, es, fr, …) works too: the agent translates the UI labels into config.json the first time.
• To change later: edit config.json and re-run node scripts/build-data.mjs. (Existing sessions keep the language they were written in; new ones use the new language.)

The dashboard

• Double-click index.html — opens via file://, no server.
• Sessions are listed in the sidebar; projects are grouped into categories with tags.
• Per-card buttons mark a project studied or hidden (stored in your browser's localStorage, so your progress is private and local).

Tag legend

Tag	Meaning
🔥	Trending right now (recent momentum, discussion, integrations)
🆕	Created recently
🏆	Top-tier — what leading practitioners actually use
📜	Still works, but has clearer successors / slowing maintenance

Tags are language-agnostic: the emoji is the stable key, only the label is translated.

---

Project structure

github-pick/
├── AGENTS.md                 # entry point any agent reads (Codex/Cursor/Cline/Aider…)
├── CLAUDE.md                 # makes Claude Code read AGENTS.md
├── workflow/github-pick.md   # canonical, agent-neutral workflow spec (source of truth)
├── config.json               # { lang, ui } — the one-time language choice
├── index.html                # the dashboard (double-click to open)
├── data.js                   # generated — do not hand-edit
├── scripts/build-data.mjs    # zero-dependency: sessions/*.json + config → data.js
├── sessions/                 # one <date>-<slug>.json (+ .md) per topic run
└── LICENSE

How the data flows

sessions/*.json (one per topic, written by the agent) → node scripts/build-data.mjs inlines config.json + all sessions into → data.js → index.html reads it. The build script is the only thing that writes data.js; agents never hand-edit it.

Works with any agent

The "brain" is a plain spec (workflow/github-pick.md) that any capable agent can follow. Agents differ only in how they dispatch the work — one that can spawn parallel subagents (e.g. Claude Code) fans the category reviews out for speed; a single-thread agent (e.g. Codex) runs them sequentially. The output is identical either way.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. The workflow logic is centralized in workflow/github-pick.md; change behavior there. Keep data.js out of manual edits; regenerate it with the build script.

License

MIT