Open-Source Newsletter Auto-Generation System: A No-Code AI Deployment Breakdown

An Information Newsletter System That AI Can Deploy For You

For professionals who need to compile large volumes of news and produce newsletters every day, manually searching, categorizing, scoring, deduplicating, and formatting is an extremely tedious task. Information sources are scattered across Weibo, Douyin, Bilibili, and various overseas websites—there's simply no way for one person to keep up. Even after collecting everything, you still have to categorize by section, sort by score, filter out duplicates, and finally write it all up in a standard format.

Bilibili creator Lao Wang recently shared an open-source solution called the "Information Newsletter Auto-Generation System." Its biggest selling point is this: you don't need to know any technology at all—you just have AI read a project document, and it will handle the entire deployment for you. The system is open-sourced on GitHub, requires zero manual operation throughout, and pushes newsletters to Feishu automatically every day. This article breaks down its technical architecture and design philosophy.

Six-Stage Automated Pipeline Design

At its core, this information newsletter auto-generation system is an information processing pipeline executed automatically by AI. The entire process is split into six interlocking stages:

• Pre-check: Automatically verifies that all tools and configurations are ready;
• Collection: Fetches content from various platforms simultaneously through four search channels;
• Classification: Automatically scores and sorts items using preset section keywords;
• Evaluation: Scores each item for popularity and filters out duplicate content;
• Generation: Automatically writes newsletter copy using preset templates;
• Publishing: Automatically creates Feishu documents, writes content into sections, and sends notifications to groups.

This pipeline-style design breaks down what was originally a complex task requiring repeated manual intervention into standardized steps that connect automatically. Throughout the entire process, the user needs to do nothing—this is a classic AI Agent workflow approach.

The "AI Agent workflow" mentioned repeatedly in the article is the key concept for understanding this system. Traditional AI applications are mostly "question-and-answer" conversational tools, whereas an AI Agent emphasizes giving large language models the ability to plan, invoke tools, and execute multi-step tasks. A typical Agent breaks down a complex goal into subtasks, autonomously decides which tools to call (such as search, code execution, file read/write), and adjusts subsequent actions based on intermediate results. Over the past two years, as the Tool Use capabilities of models like GPT-4 and Claude have matured, and as local AI Agent tools like Cursor and Claude Code have become widespread, it has become possible for AI to directly read project documents and set up environments hands-on. This newsletter system is precisely designed as a "six-stage pipeline" that an Agent can execute step by step—essentially solidifying a human engineer's deployment experience into operational instructions that AI can understand.

Four-Channel Search: Covering 16+ Domestic and International Platforms

The system's most central capability is search. It's designed with four parallel multi-platform search channels, aiming to cover as broad a range of information sources as possible:

Channel One: Platform Built-in Search

Directly calls the results of mainstream search engines. It has the broadest coverage and requires no additional configuration.

Channel Two: Agent Reach

This is an MIT-licensed open-source multi-platform search tool that covers 16 platforms including Bilibili, YouTube, Weibo, and Zhihu. Seven of these platforms work right after installation, while several others (such as Douyin, Twitter, and Xiaohongshu) require login credentials to be configured.

Channel Three: Media Crawler

Specializes in crawling social media content from Weibo and Douyin. After scanning a QR code to log in the first time, it runs fully automatically.

Channel Four: RSS Subscription

Subscribes to various international and industry-specialized websites via Feed Parser, automatically fetching the latest content.

Worth mentioning is that the RSS subscription and Feed Parser used in the fourth channel represent a long-standing yet still practical technology. RSS (Really Simple Syndication) is an XML-based content distribution format. By publishing an RSS feed, websites allow users to get the latest updates without visiting each site individually. Although RSS was once considered "outdated" after Google Reader shut down in 2013, its value in information aggregation and decentralized subscription has never disappeared—it's especially suitable for crawling specialized websites and international media that don't have open APIs. Feed Parser is a library for parsing these RSS feeds, converting XML structured data into program-processable items. Compared to calling commercial search APIs, the RSS approach is completely free and stable, which is one of the important reasons this entire system can achieve "zero cost."

With all four channels working simultaneously, a single run can cover multiple types of information sources including social media and specialized websites, solving the pain point of "flipping through one platform after another."

Keyword Scoring: Intelligent Classification and Deduplication Mechanism

How does the system categorize the massive amount of collected information? It uses keyword scoring to achieve automatic classification. Each section is equipped with a set of classification keywords. Taking a tech news newsletter as an example, you might set up sections like "Industry Trends," "Hot Topics," "In-Depth Analysis," "Cutting-Edge Technology," and "Policy Interpretation."

The classification logic is straightforward: when each item comes in, the system checks which keywords it hits—hitting a section keyword adds points, hitting an exclusion keyword deducts points, and the section with the highest score becomes its category.

After classification, the process enters the evaluation stage. The system assigns a popularity score to each item, downweights those with low confidence, and performs deduplication to ensure no repeated content appears in the newsletter. Each section is sorted by popularity, and the Top 5 are selected—both the quantity and threshold can be adjusted in the configuration.

Configuration-Driven: Switch Domains by Changing Just One File

The most noteworthy design of this system is its "configuration-driven" approach. All section keywords, search terms, subscription sources, and writing style preferences are managed centrally in a single configuration file.

This means that if you want to switch from "tech news" to the "finance domain," you don't need to modify any code—you just replace the keywords and section definitions with finance-related content. The system's six-stage pipeline, four-channel search, scoring and deduplication logic, and Feishu publishing process are all reused. Want to do entertainment gossip? Same operation.

What's even more convenient is that this configuration file doesn't need to be written by hand either—after AI finishes reading the project document, it automatically generates the initial configuration. The user just needs to specify what domain they want to cover. This design greatly reduces migration costs and is key to the system's versatility.

Copy Generation and Automatic Feishu Publishing

Copy generation uses formulaic templates. Each newsletter piece contains four parts: Value Point (core information, no more than 80 characters), Title (no more than 60 characters), Body (300-400 characters), and Source Information with the original link.

Each section also has its own style preferences. For example, "Industry Trends" uses an objective and neutral tone focusing on facts, while "Opinion & Commentary" uses an analytical tone focusing on distilling viewpoints. These are all preset in the configuration file and applied automatically during generation. Users can also send their own commonly used templates to the AI for learning, replacing the default format.

Once generation is complete, the system automatically creates a Feishu cloud document and writes content into sections, with each piece accompanied by a source link, and finally sends a notification to the Feishu group. Users can view and forward it simply by opening Feishu on their phone.

The system's choice of Feishu (Lark) as the final publishing channel has its technical and ecosystem considerations. Feishu, launched by ByteDance, offers relatively comprehensive open platform capabilities, including custom bots (Webhooks), cloud document APIs, group message push interfaces, and more—all free for personal use. For an automation system, this means it can programmatically create structured documents, write content into sections, and trigger group notifications—all without human intervention. Compared to email or plain-text push, Feishu cloud documents support rich text formatting and multi-device syncing. Users can view and forward on mobile, providing an experience closer to a finished newsletter. This practice of combining open APIs with collaboration tools is becoming a common pattern for the implementation of personal automation projects.

A Zero-Technical-Barrier AI Deployment Method

The biggest difference between this system and other tools lies in its deployment method. Users don't need to install anything or write any code. The only thing they need to do is have AI read a document in the project called AGENTS.md.

After reading it, the AI will guide the user step by step to provide necessary information (such as Feishu bot configuration, Weibo and Douyin logins, etc.) and automatically complete: checking the computer environment, installing all components, deploying the four search channels, configuring the Feishu connection, creating scheduled tasks, and testing and verifying. The user just needs to click and scan QR codes as prompted.

After deployment, the system sets up a daily scheduled task that automatically runs the entire process at the appointed time. The system also accumulates deduplication records, becoming more accurate the more it's used, and search results are cached to improve efficiency.

Cost and Target Users

It's worth emphasizing that this system is completely free. Among the four search channels, the platform built-in search is free, and Agent Reach (MIT open-source), Media Crawler, the Feed Parser used for RSS, the open-source tools used for Feishu publishing, and the personal version of Feishu are all free. Classification, scoring, deduplication, and copy generation are all built into the system and don't call any paid interfaces. Adding up all the components, the total cost is zero. The only requirement is a computer plus a local AI Agent tool.

As for target users, anyone who needs to regularly collect news can benefit:

• Information reporting workers: Regularly compile news and produce newsletters in a fixed format;
• Content operations personnel: Continuously track industry hotspots and generate competitive analysis;
• Independent content creators: Need abundant sources of inspiration and hotspot tracking;
• Team managers: Stay informed about industry trends and competitive intelligence daily.

Conclusion

This system embodies a typical trend in current AI Agent applications: using natural-language-driven AI to complete the deployment and operation of complex tool chains. Its value lies not only in being free and open-source, but more importantly in combining "configuration-driven" design with "AI automatic deployment," enabling non-technical users to have an automated information processing pipeline.

Of course, systems like this still depend on the stability of login states across platforms and the availability of search tools, so actual results need to be verified through long-term use. But for professionals troubled by information compilation work every day, this undoubtedly offers a low-barrier solution worth trying. The project is already open-sourced on GitHub, and interested readers can obtain it themselves.

Key Takeaways