shield_lock
JinDaGe
← Back to skills
extension
Category: Data & AnalyticsNo API key required

每日 AI 资讯简报

一个每日 AI 资讯聚合与推送技能：用 agent-browser 从 AIbase（aibase.com/zh/news）和量子位（qbitai.com）实时抓取当日最新 AI 头条，整理成白底黑字、带来源与时间的 HTML 简报，并通过 present_files 推送。当用户说"每日AI资讯""AI日报""AI早报""抓取AI资讯""AI新闻推送"或想配置每日定时 AI 资讯推送时使用。

personAuthor: user_8e21960chubcommunity
downloadDownload package

每日 AI 资讯简报 (Daily AI News Digest)

Overview

一个经过真实生产验证的端到端工作流：用 agent-browser CLI 从中国科技媒体站点抓取当日最新 AI 头条，整理成白底黑字的 HTML 简报。该技能已经过实际运行检验，并规避了裸用浏览器自动化时常见的环境坑（Chromium 下载超时、链式命令偶发失败、自动化时区与推送机制等）。

When to use

  • User asks for a daily / morning AI news digest ("每日AI资讯", "AI日报", "AI早报", "抓AI新闻").
  • User wants AI news aggregated from AIbase and 量子位 (QbitAI).
  • User wants to set up a recurring scheduled AI-news push.

权限与安全声明（平台审核参考）

本技能完全在本地运行，不收集、不上传任何用户隐私数据。其涉及系统资源的行为如下，已在脚本与提示词中明确约束：

  • 启动本地浏览器：通过 agent-browser CLI 拉起本机已缓存的 Chrome（仅用于渲染新闻网页、读取正文），不会访问用户其他文件或网页。
  • 网络下载（仅一次）：scripts/setup_browser.sh 仅从 npmmirror 官方 CDN（cdn.npmmirror.com）下载指定版本的 Chrome-for-Testing 压缩包并缓存到用户目录；无硬编码密钥、无第三方数据外传。若机器已有可用 Chrome，可跳过此步、直接设置 AGENT_BROWSER_EXECUTABLE_PATH。
  • 读写工作区：仅在本技能所在工作目录生成 每日AI资讯_YYYY-MM-DD.html 简报文件，不修改用户其他文件。
  • 无外部 API Key / Token：不调用任何需要密钥的付费接口。

安全提示：脚本内唯一的下载源是 npmmirror 官方二进制镜像，请审核时据此确认无恶意网络行为。

Prerequisites (one-time per machine)

  1. Install the CLI: npm install -g agent-browser.
  2. Provide a Chromium binary. In restricted networks the built-in agent-browser install (downloads from Google) often times out. Run scripts/setup_browser.sh to fetch the same Chrome-for-Testing build from the npmmirror China CDN and cache it at C:/Users/Administrator/.agent-browser/chrome/. The script is idempotent.
  3. During every run, export these two variables (the Rust binary wants a Windows-style path, not POSIX): 
    export AGENT_BROWSER_EXECUTABLE_PATH="C:/Users/Administrator/.agent-browser/chrome/chrome-win64/chrome.exe"
export AGENT_BROWSER_ARGS="--no-sandbox"

Workflow

Step 1 — Fetch sources with agent-browser

agent-browser keeps a persistent background daemon: the first open starts it, later commands reuse it, close ends it. Run each major step as its own shell call — chained ; commands occasionally fail on the first open (see pitfalls).

export AGENT_BROWSER_EXECUTABLE_PATH="C:/Users/Administrator/.agent-browser/chrome/chrome-win64/chrome.exe"
export AGENT_BROWSER_ARGS="--no-sandbox"

agent-browser open "https://www.aibase.com/zh/news"
agent-browser wait --load networkidle
agent-browser snapshot > /tmp/aibase.txt

Then the second source as a separate call:

agent-browser open "https://www.qbitai.com/"
agent-browser wait --load networkidle
agent-browser snapshot > /tmp/qbitai.txt
agent-browser close

Extract headlines + timestamps from the two text snapshots:

  • AIbase: lines containing 阅读文章 (each entry also carries a relative time like 2小时前 / 1天前).
  • 量子位: heading "..." lines are the article titles.

Step 2 — Compile the white-theme HTML digest

Copy assets/digest-template.html as the starting point — it already carries the mandatory white-background / black-text :root palette plus light source/time tags. Replace the placeholder content with the day's real news, grouped into sections (typical: 大模型与公司动态 / 产品与发布 / 安全与法律 / 机器人与硬件 / 研究与仿真). Each item records 来源 + 时间. Save as 每日AI资讯_YYYY-MM-DD.html.

Hard rules for the output:

  • Never write a dark default. Never rely on @media (prefers-color-scheme: dark).
  • Keep the light palette as the only mode (any device/preview shows white bg + black text).
  • Source tags: light-blue background, blue text. Time tags: light-gray background, gray text. No dark/near-black chip backgrounds.

Step 3 — Present the result

Use present_files on the generated HTML so the user can open it on desktop and mobile.

Step 4 (optional) — Schedule as a recurring automation

Create a daily automation. Critical settings (details in references/troubleshooting.md):

  • rrule MUST include TZID=Asia/Shanghai, e.g. FREQ=DAILY;BYHOUR=9;BYMINUTE=0;TZID=Asia/Shanghai.
  • Embed the browser env vars and the rule "do not run agent-browser install; reuse the cached Chrome" into the automation prompt.
  • Reality of delivery: runs land in PENDING_REVIEW and do not auto-push a card; the client must be online around the trigger time, or the day is silently skipped (no backfill). The most reliable "push" is to re-present the generated HTML via present_files when the user asks.

Resources

  • icon.png — 512×512 市场展示图标（蓝色新闻卡片风格）。
  • assets/digest-template.html — white-background / black-text report template (start here).
  • references/troubleshooting.md — environment pitfalls and automation delivery notes.
  • scripts/setup_browser.sh — idempotent Chromium cache setup from npmmirror.

Key pitfalls (summary)

  • Chained ; commands: the first open sometimes exits 1 with no output. Run open / wait / snapshot as separate Bash calls; re-running usually succeeds.
  • Stuck daemon: if a fetch hangs, taskkill /F /IM chrome.exe and taskkill /F /IM agent-browser.exe, then retry.
  • Chromium download from Google times out → use the npmmirror cached build via scripts/setup_browser.sh.
  • close connection error: the snapshot is usually already saved before close runs; the error is harmless to ignore.