chengjialu8888/ai_news_digest

---

中文

AI News Digest 是一个面向 AI 从业者、投资研究者、产品经理、内容创作者和开发者 的日报生成引擎。它不是再做一个“AI 总结器”，而是把散落在中英文媒体、官方博客、GitHub、Hacker News、Hugging Face、X Builder 动态和 RSS Feed 里的信号，整理成一份可以直接阅读、追踪和二次加工的行业情报。

你可以用它每天生成：

• 一份 3 分钟可读完 的 AI 行业日报
• 每条新闻都带 原始链接 的可追溯摘要
• 微信公众号文章可抓取 全文 Markdown 归档，不只依赖搜索片段
• 带信号等级的内容筛选：重磅 / 值得关注 / 常规
• 可沉淀到表格、看板或知识库的 12 列 CSV
• 飞书 Newsrun 卡片、详细版文档、结构化沉淀文档、CSV 和 HTML 五类产物
• 每天从 BidClub.ai 扫描过去 24 小时新增播客，把节目观点作为「海外建设者」的补充入口，并保留原始收听链接与有限证据摘录
• 详细版默认配套由 imagegen 直接生成“图 + 字”的三张 16:9 趋势 PNG；需要艺术化输出时，可选开启艺术家 Lottery：从 MoMA、Centre Pompidou 等现代/当代机构，以及 The Met、大英博物馆、卢浮宫、雅典国立考古博物馆、柏林国家博物馆群/佩加蒙博物馆、墨西哥国立人类学博物馆和哥伦比亚国家博物馆的官方目录抽取艺术家、艺术运动、工坊传统或文化视觉系统。开启后每日风格会追加到统一的「AI日报｜艺术风格 Lottery」飞书文档；HTML 仍是独立日报阅读产物，不是趋势图默认中间格式
• 初创动向 每日优先扫描 AI 产品上游供给层与下游产品层的 AI-native startup，关注现金流/经营信号和成长线索；普通垂类公司不纳入，不把融资额当作商业成功
• 适合团队同步、选题会、投研记录和公众号素材库的结构化输出
• 面向长程运行的 Codex Goal × Ralph Loop × Kleisli Gate 工作流
• 首次使用前先询问你的产物偏好，避免把个人速读、团队同步、投研和内容运营混成同一种交付

它解决什么问题？

长程运行架构

AI News Digest 现在不只是一个“日报提示词”，而是一个可以持续跑完复杂信息工作的分阶段系统。

适合谁？

一眼看懂：工作流与产品架构

工作流

flowchart LR
    A["60+ 中英文信源 + BidClub 播客"] --> B["抓取与反爬降级"]
    B --> C["去重与交叉验证"]
    C --> D["信号分级"]
    D --> E["事实核验"]
    E --> F["Markdown 日报"]
    E --> G["12 列 CSV"]
    F --> H["询问用户补充"]
    G --> H
    H --> I["最终版 Markdown 三大趋势"]
    I --> J["ImageGen 三张 16:9 图 + QA"]
    J --> K["飞书详细版 + 结构化沉淀"]
    K --> L["Newsrun 卡片推送"]

产品架构

Trust Layer

AI News Digest 的核心价值不是“写得更像日报”，而是让日报更可信。

质量检测环节的设计灵感来自 Signex：把分散信号先沉淀为可检查的数据，再通过 source health、去重、分析视角和反馈记忆，让日报不只是“聚合”，而是逐步变成一个可校准的情报工作流。

Gate 1: 数据源健康检查  -> 信源是否正常返回、内容是否足够新
Gate 2: 去重与交叉验证  -> 合并同一事件，优先保留一手来源
Gate 3: 信号分级        -> 标记重磅、值得关注、常规信息
Gate 4: 事实核验        -> 检查公司名、时间线、数字、引用和链接
Gate 5: 完整性自检      -> 确认日报板块覆盖完整，无明显遗漏

Source System

AI News Digest 使用 9 层信源系统，把“覆盖面”和“可靠性”拆开处理。

除了 follow-builders 的固定播客 Feed，日报每天还会检查 BidClub.ai 的最新单集分页。它把发现、详情和事实核验拆成不同阶段：阶段文件只保留有限证据摘录，关键判断回到原始节目链接，避免长 transcript 撑大上下文。

Outputs

第一次运行前，工作流会先询问产物偏好：使用场景、产物组合、语言语气、推送位置和沉淀规则。确认后写入 data/output-preferences.json，后续运行默认复用；这让同一套日报可以稳定服务个人速读、团队晨会、投研记录、内容运营和知识库沉淀。

各类产物各自承担不同工作：

可选艺术家 Lottery 视觉后端

默认视觉后端是 imagegen；需要使用艺术家/艺术运动作为当日视觉语法时，显式开启：

VISUAL_BACKEND=artist-lottery \
  python ralph-daily-loop/stage9.py

Lottery 链路会按日期可复现地从全球博物馆目录抽签，现代/当代与古典/跨文明并置；Art Basel 各地官方名录仍只作为现代/当代发现层。它会把 style_lottery、博物馆/艺术博览会来源、时期、地域、目录查询词、三条趋势适配、三张 16:9 PNG、表格布局和 QA 结果写入 output/daily-trends.json 与 output/art-style-lottery-entry.json；只有开启此后端时，才追加统一的「AI日报｜艺术风格 Lottery」飞书文档。可用 VISUAL_STYLE_OVERRIDE 指定目录内风格，RECENT_VISUAL_STYLES 排除近期风格。

Lottery 2.0：从随机风格到可积累的视觉研究

每次抽签现在生成一份可执行的 style_contract，把馆藏 provenance 翻译成六段 prompt 组件：机制事实、趋势命题、主体/空间关系、材料/光、信息层级、排除项。evaluation_contract 把 QA 失败定向返回到 catalog、mechanism_extract、trend_translation、prompt 或 craft，而不是笼统地写“再优化一次”。评估器只看最终截图、趋势输入和契约，不读取生成器私有推理或自评。

审核结果可用 visualization/artist-lottery/record_case_review.py 追加到 data/art-style-cases.json。下一次 Lottery 会参考已审核的通过、返工和拒绝案例，避免机械重复，并保留 case_id、失败类型、修复动作和用户反馈。案例库让艺术风格从每日皮肤变成可以校准的长期审美资产，也让 prompt 质量、视觉质量和趋势适配逐步形成自己的数据集。

QA 采用 Qwen-Image-Bench 的五个一级维度：Quality、Aesthetics、Alignment、Real-world Fidelity、Creative Generation；日报工作流只做人工参考评审，不声称运行 Q-Judger。维度与评分参考

Design prompt method：把审美变成可积累的工作方法

视觉输出现在统一遵守 DESIGN_PRINCIPLES.md 的四阶段方法：discover 先探索至少三个有结构差异的方向，define 定义视觉身份与信息不变量，enrich 用 ImageGen 建立主体/材料/状态，deliver 让新上下文的独立 critic 审核最终截图并完成删减审计。它同时作用于默认 ImageGen 和 Artist Lottery，因此 Lottery 不再是随机挑一个名字，ImageGen 也不再是直接把趋势套进一个漂亮模板。

每次运行会保存 seed fingerprint、方向候选、未采用 prompt、选择理由、critic 评分、删除记录和 QA 回流动作；艺术家姓名仍只作 provenance，参考图只作 moodboard。长期价值是把“我喜欢/我不喜欢”变成可复盘的审美数据：下一次运行可以避开机械重复，编辑可以解释每个视觉决定如何服务趋势，运营可以稳定地产出适合不同渠道的资产，而不是每次从上下文里重新猜风格。

Mck PPT Design 趋势可视化

仓库已接入 likaku/Mck-ppt-design-skill（当前远端版本：e190e083）。它作为用户明确指定时使用的咨询式视觉后端，不替换默认的 imagegen 图像生成：同一份 daily-trends.json 可以按场景输出飞书插图或三页 PPT；趋势可视化 HTML 不默认生成。

Goal 9 会在 output/daily-trends.json 中保存最终三条趋势和佐证。确认最终 Markdown 后运行：

python visualization/mck-ppt-design/render_daily_trends.py \
  --trends-json output/daily-trends.json \
  --out-dir output/mck-ppt-design

渲染器遵循 Mck 的 S3/S4 机器门禁，产出 content.json、gate_s3.json、gate_result.json、三页 .pptx 和 trend-1-mck.png 至 trend-3-mck.png。PNG 固定为 16:9，可直接插入飞书详细版；如需公众号 HTML，必须由用户明确指定。完整来源仍留在日报正文，视觉层只承载判断入口。

HTML Editorial 趋势可视化

工作区最新视觉规则已沉淀为一个显式可选后端：html-editorial。它不是默认日报 HTML 的替代品，而是面向浏览器、HTMLbox、飞书截图和 16:9 section 的趋势阅读资产。它继承 DESIGN_PRINCIPLES.md、日报图文一体架构 和 品牌视觉契约：判断先于装饰，艺术机制与品牌层分离，三屏构图必须有结构差异，证据边界必须可见，移动端正文必须可读。

VISUAL_BACKEND=html-editorial \
  python visualization/html-editorial/render_daily_trends.py \
  --trends-json output/daily-trends.json \
  --out-dir output/html-editorial

它会生成 index.html、trend-1.png 至 trend-3.png、contact-sheet.png 和 qa.json。完整来源仍回到日报正文；HTML 后端只消费 Goal 9 已确认的 daily-trends.json，不重新研究或扩大事实。详细规则和后端比较见 visualization/html-editorial/README.md。

Markdown / 飞书详细版文档

# AI 日报 2026-04-07

## 一句话总结
> 今天的主线：模型能力、Agent 工具链和 AI 基础设施继续加速。

## 偏 fact 类新闻

### 大厂动向
1. [重磅] OpenAI 发布新产品更新 -- 摘要... [[Source]](https://example.com)

### 初创 / 融资
### 生态 / 政策

## 偏观点类
## 海外建设者动态
## 质量审核报告

CSV 结构化数据

本地输出路径

output/feishu-card.md
output/daily-report.md
output/structured-archive.md
output/daily-report.csv
output/daily-report.html

快速开始

Claude Code

git clone https://github.com/chengjialu8888/AI_News_Digest.git ~/.claude/skills/ai-news-digest

Codex

git clone https://github.com/chengjialu8888/AI_News_Digest.git ~/.codex/skills/ai-news-digest

OpenClaw

git clone https://github.com/chengjialu8888/AI_News_Digest.git ~/skills/ai-news-digest

Mira

cp -r ai-daily-report /opt/tiger/mira_nas/plugins/prod//skills/

其他个人助手

curl -L https://raw.githubusercontent.com/chengjialu8888/AI_News_Digest/main/ai-daily-report/SKILL.md -o ai-news-digest.system.md

安装后对你的助手说：“跑一下今天的 AI 日报”或“出一期适合团队晨会的 AI 行业日报”。首次运行时，它会先问你产物偏好；通用个人助手可把 ai-news-digest.system.md 作为 System Prompt 使用。

项目结构

AI_News_Digest/
├── ai-daily-report/
│   ├── SKILL.md
│   ├── evals/
│   └── examples/
├── news-aggregator-skill/
├── smart-web-fetch/
├── content-trend-researcher/
├── wechat-article-fetch/
├── visualization/
│   ├── artist-lottery/
│   │   ├── README.md
│   │   └── record_case_review.py
│   ├── mck-ppt-design/
│   └── html-editorial/
├── ralph-daily-loop/
│   ├── style_contracts.py
│   ├── fetch_bidclub.py
│   ├── stage9.py
│   └── stage9_kleisli.py
├── data/
│   └── art-style-cases.example.json
├── assets/
│   └── github-header.svg
├── ai-daily-report.skill
└── README.md

Evaluation

---

English

AI News Digest is a daily intelligence engine for AI operators, researchers, investors, builders, and content teams. It turns scattered signals from Chinese and English media, official blogs, GitHub, Hacker News, Hugging Face, X builder feeds, and RSS sources into a verified daily brief.

It helps you produce:

• A readable AI industry brief in about 3 minutes
• Source-linked summaries for every news item
• Full Markdown archives for WeChat articles, not just search snippets
• Signal grading for what is important, interesting, or routine
• A 12-column CSV for databases, spreadsheets, dashboards, and knowledge bases
• Five outputs: Feishu Newsrun cards, detailed docs, structured archive docs, CSV, and HTML
• A reusable workflow for team syncs, research notes, editorial planning, and market tracking
• A long-running Codex Goal × Ralph Loop × Kleisli Gate workflow for complex daily intelligence
• A first-run preference check so personal scanning, team distribution, research, and content operations do not get forced into the same output shape

Why Star This Repo?

Star this repo if you want a practical starting point for building a reliable AI news workflow:

• It focuses on source-backed intelligence, not generic summaries.
• It combines Chinese context + international first-hand sources.
• It can fetch mp.weixin.qq.com articles into local Markdown archives for traceable Chinese-source verification.
• It ships with a reusable Mira Skill / system prompt.
• It outputs Feishu cards, detailed docs, structured archives, CSV, and HTML.
• It is designed to keep running as new AI sources, MCP servers, feeds, newsletters, and QA gates appear.
• It treats long-context failure as a product problem: stages, checkpoints, files, and traces are part of the design.
• It turns daily news into an operating system for memory: each issue becomes a readable brief, a pushable card, and a queryable dataset for future trend reviews.

Tier 5 builder signals come from follow-builders, which tracks 25 curated AI builders on X, 6 AI podcasts, and official Anthropic/Claude blog updates.

The daily scan also checks BidClub.ai for new podcast episodes and recorded conversations. It keeps discovery metadata and a bounded evidence excerpt in the stage snapshot, while the original listening link remains the verification source.

The QA workflow is inspired by Signex: source health checks, signal convergence, analysis lenses, and feedback memory are adapted here for daily AI news production.

Long-Run Architecture

AI News Digest is no longer just a prompt for a daily brief. It is a staged system for finishing a large, failure-prone intelligence job without blowing up the context window.

Quick Start

Claude Code

git clone https://github.com/chengjialu8888/AI_News_Digest.git ~/.claude/skills/ai-news-digest

Codex

git clone https://github.com/chengjialu8888/AI_News_Digest.git ~/.codex/skills/ai-news-digest

OpenClaw

git clone https://github.com/chengjialu8888/AI_News_Digest.git ~/skills/ai-news-digest

Mira

cp -r ai-daily-report /opt/tiger/mira_nas/plugins/prod//skills/

Other Personal Assistants

curl -L https://raw.githubusercontent.com/chengjialu8888/AI_News_Digest/main/ai-daily-report/SKILL.md -o ai-news-digest.system.md

After installation, ask your assistant: “Generate today’s AI daily report” or “Create an AI industry brief for my team sync.” On the first run, it will ask for output preferences before producing or pushing artifacts. For generic assistants, use ai-news-digest.system.md as the system prompt.

Output Workflow

The workflow produces a set of operational artifacts and asks for preferences before the first run:

Artist Lottery as a cumulative visual research system

The Lottery backend now emits a style_contract, not just an artist name. It turns museum provenance into six executable prompt slots: mechanism fact, trend proposition, subject/spatial relation, material/light, information hierarchy, and exclusions. An evaluation_contract routes a failed review back to catalog, mechanism_extract, trend_translation, prompt, or craft, so “make it better” becomes a concrete repair action. The evaluator receives the delivered image, trend input, and contract, not the generator’s private reasoning or self-assessment.

Review results can be appended with visualization/artist-lottery/record_case_review.py to data/art-style-cases.json. Future draws use reviewed accepted, rework, and rejected cases to reduce mechanical repetition and retain case_id, failure type, repair action, and user feedback. This turns visual taste into a calibratable operating asset for long-run editorial and knowledge work rather than a disposable daily skin.

Mck PPT Design trend visuals

The repository now includes likaku/Mck-ppt-design-skill (remote revision e190e083) as an optional consultant-style visual backend. The default visual backend is imagegen; the museum-grounded artist Lottery backend is also opt-in. An additional html-editorial backend is available when the user explicitly wants browser, HTMLbox, Feishu screenshot, or three 16:9 section assets.

Goal 9 writes the final three trends and their evidence to output/daily-trends.json. After the Markdown is confirmed, run:

python visualization/mck-ppt-design/render_daily_trends.py \
  --trends-json output/daily-trends.json \
  --out-dir output/mck-ppt-design

The renderer follows Mck’s machine-readable S3/S4 gates and produces content.json, gate_s3.json, gate_result.json, a three-slide .pptx, and trend-1-mck.png through trend-3-mck.png. The PNGs are fixed at 16:9 for Feishu; WeChat HTML requires an explicit user request. Full source evidence remains in the detailed report while the visual layer provides the decision entry point.

HTML Editorial trend visuals

The latest workspace visual rules are available as an explicit html-editorial backend. It renders the same bounded daily-trends.json into a self-contained HTML reading system: three structurally different 16:9 frames, three PNG screenshots, and a contact sheet. The frames use a double-entry boundary, an open loop with a verification break, and split task/evidence windows. Each screen keeps one judgment, 2-3 mechanism/evidence nodes, and one limitation or next observation. It is designed for browser preview, HTMLbox, Feishu screenshots, and editorial review; it does not replace the normal daily-report HTML.

VISUAL_BACKEND=html-editorial \
  python visualization/html-editorial/render_daily_trends.py \
  --trends-json output/daily-trends.json \
  --out-dir output/html-editorial

The backend inherits the repository’s brand contract, integrated type-and-image rules, mobile readability checks, and evidence boundaries. It consumes Goal 9 output only; full source links and detailed facts stay in the report rather than being invented inside the visual layer.

Design prompt method: a durable taste operating system

All visual backends follow the repository-level DESIGN_PRINCIPLES.md contract: discover at least three structurally different directions, define a concrete visual identity and information invariants, enrich the world with ImageGen, and deliver through a fresh-context critic plus a deletion audit. This applies to default ImageGen as well as Artist Lottery. Lottery is therefore not a random artist-name picker, and ImageGen is not a one-shot decoration pass.

Each run persists a seed fingerprint, direction candidates, discarded prompts, selection rationale, critic scores, deletion decisions, and QA return routes. Artist names remain provenance and references remain moodboards, not copy targets. The long-run value is operational: taste becomes reviewable data, future runs avoid mechanical repetition, editors can explain why a visual clarifies a trend, and the team can keep producing channel-specific assets without reconstructing the design brief from a crowded context window.

How It Compares

License

MIT

---

Note: The example report title, date, and https://example.com link above are placeholders used to show the output format. Replace them with real daily report data before publishing generated examples.