Skip to content

skill.md

Latest commit

 

History

History
103 lines (79 loc) · 5.56 KB

skill.md

File metadata and controls

103 lines (79 loc) · 5.56 KB
Raw

AI News Weekly Agent

Generate a weekly AI report in Markdown, including official announcements, industry news, GitHub-oriented open-source signals (via configured feeds), a capped paper ratio, and optionally OpenClaw top skills as auxiliary context.

When to use

  • User asks for weekly AI digest/report
  • User wants GitHub / open-source release or repo activity summarized together with general AI news
  • User wants OpenClaw skills ranking in addition to the above (not as the sole “source of truth” for tooling facts)

Weekly layout & synthesis (newsletter-style)

Output follows a fixed Markdown template generated by the pipeline:

  1. 小结：3–4 句总览（本周主线、重要变化）
  2. 正文：按 sources.json 分类的逐条资讯（含链接、日期、LLM 中文解读）
  3. OpenClaw 技能热榜（仅当启用）：出现在正文条目之后，标题标明「社区热度 · 辅助参考」
  4. OpenClaw 链接 / 信源与原文链接 / 生成信息

宿主 Agent 在编排用户意图时，应把「周刊」理解为：先事实与可核验链接，再解读与价值；不要跳过信源配置去编造条目。

Research discipline (deep-research–inspired)

These rules apply to how the hosting agent describes tasks and how the pipeline uses LLM enrichment (see project/prompts/digest_llm_prompts.py):

  • Prefer verifiable items: official blogs, RSS, and GitHub Release / repo feeds present in sources.json
  • Treat OpenClaw 热榜 as a popularity snapshot, not a substitute for Release or repo facts
  • When the user gives intent_text, keep queries concrete (repos, products, time window) so GNews / Playwright supplements stay on-topic

GitHub vs OpenClaw（主辅关系）

角色 含义
主（GitHub / 仓库事实） github.com 相关条目（如 Release、仓库公告、*.atom commits）：叙事与小结中的「开源/工具链」应优先依据这些可点击事实。
辅（OpenClaw） topclawhubskills.com 榜单：用于对照社区在用什么技能；不得用排名替代 GitHub/官方发布中的版本与变更事实。

宿主说明：若用户同时关心「GitHub 动态」和「技能榜」，仍走默认采集 + 可选 OpenClaw；小结里应先概括仓库/官方侧事实，再用一两句提及热榜与主线的呼应（具体文风由 enrich 提示约束）。

Skill routing (project agent)

  • Default route: run the news digest pipeline only (sources.json + site crawlers + optional GNews as configured).
  • OpenClaw route: enable when user intent explicitly mentions keywords such as openclaw, topclaw, 技能榜, 榜单, skill, or when the user clearly wants leaderboard data alongside the weekly news (not instead of it).
  • When OpenClaw is active, the pipeline fetches the leaderboard and injects it after the main news body; LLM enrichment receives instructions that the leaderboard is auxiliary.
  • If the user does not want OpenClaw ranking, do not fetch or render the OpenClaw section.

GitHub-oriented intent (e.g. github, release, 开源, 仓库): prefer emphasizing configured GitHub feeds in sources.json; do not auto-enable OpenClaw unless the user also asks for 榜单 / OpenClaw.

Command

python3 run_daily_digest.py --use-llm --llm-provider auto --window-hours 168 --max-paper-ratio 0.2 --min-official-items 3

Required environment variables

  • Ark mode:
    • ARK_API_KEY
    • ARK_MODEL
    • ARK_ENDPOINT_ID (recommended, ep-xxx)
  • OpenAI-compatible mode:
    • OPENAI_API_KEY
    • OPENAI_BASE_URL (if ends with /v1, script auto-expands to /chat/completions)
    • OPENAI_MODEL (optional, but recommended)

Notes:

  • In OpenClaw, these variables may be injected by the runtime.
  • If not injected, users must provide them manually in runtime env or local .env.
  • --llm-provider auto prefers Ark when Ark vars exist; otherwise uses OpenAI-compatible mode.
  • Default security policy requires HTTPS endpoint and allowlisted LLM hosts.

Optional environment variables

  • DIGEST_WEBHOOK_URL: webhook for send notifications when using --send

Environment template

Use runtime env injection in OpenClaw, or copy this template to local .env:

ARK_API_KEY=
ARK_MODEL=Doubao-Seed-1.6-lite
ARK_ENDPOINT_ID=
OPENAI_API_KEY=
OPENAI_BASE_URL=
OPENAI_MODEL=
DIGEST_WEBHOOK_URL=

External network access

This skill may access these external endpoints during execution:

  • https://ark.cn-beijing.volces.com (Ark LLM API)
  • OPENAI_BASE_URL target (OpenAI-compatible provider endpoint)
  • https://topclawhubskills.com (OpenClaw leaderboard snapshot, when enabled)
  • RSS/news URLs listed in sources.json (including GitHub *.atom feeds where configured)
  • Optional webhook hosts for notification: Feishu/Lark or DingTalk

Security and data handling

  • No API keys are hardcoded in repository files.
  • .env is not tracked and should never be committed.
  • Only report-generation inputs are sent to LLM when --use-llm is enabled.
  • Users should review and trust the configured endpoints before running.
  • SSL verification is enabled by default. Insecure SSL fallback requires explicit --allow-insecure-ssl.
  • Custom LLM endpoint hosts are blocked by default unless --allow-custom-llm-endpoint is provided.
  • Optional extension: set LLM_ALLOWED_HOSTS (comma-separated) to add trusted hosts.
  • Webhook only supports HTTPS and official Feishu/Lark/DingTalk domains.

Publishing note

  • For ClawHub publish, confirm the MIT-0 license checkbox in the web UI.