Daily Recommendation Ads and LLM Tech Brief

Problem

Public signals for recommendation systems, ads, and LLM/Agent engineering are scattered across WeChat Official Accounts, Zhihu, and GitHub. The feed is often mixed with recruiting posts, PR releases, and event announcements, making it hard to isolate real engineering progress or algorithmic changes day to day.

How It Works

The skill splits the workflow into two pipelines:
- Article intelligence: search mp.weixin.qq.com and zhihu.com using after:{{yesterday}}, focused on entities such as Tencent, Alibaba, ByteDance, Baidu, Huawei, Meituan, Zhipu AI, and DeepSeek. It separates Ads & Recommendation from LLM/Agent using terms like recall, ranking, CTR, CVR, Agent, and distributed training, then merges duplicate threads about the same topic.
- GitHub trends: scan global repositories for topic:recommendation-system, topic:llm, topic:agent, inference framework, and related terms, ranked by repository traction, forks, and update time across Daily, Weekly, and Monthly windows.

Each item should include source, title, a technical summary, and links. GitHub entries also classify the project as weights, training framework, inference engine, or agent app, and report trend data. The execution rules prioritize de-duplication, randomized search intervals, and content with architecture diagrams, formula derivation, benchmarks, or incident postmortems, while filtering out recruiting, award PR, and empty event notices.

Boundaries

This is best treated as a daily engineering brief for rec/ads and LLM/Agent, not a general news summarizer. Output quality depends on the monitored entity list, the presence of substantive technical detail, and the signal-to-noise ratio of GitHub repositories. Monthly trend analysis also depends on continuous logs in /data/monitoring_log.md.