Enterprise Sentiment Monitor

Problem

Public discussions about companies, brands, or people are scattered across news sites, social platforms, and forums. Manual daily searches easily miss aliases, duplicate reports, and escalating negative mentions. A sentiment pipeline that sends every item to an LLM also creates unnecessary token cost. This skill targets Chinese public-opinion monitoring and moves targets, keywords, recipients, schedule, and data paths into monitor_config.json, reducing hardcoding and making it easier to switch monitoring subjects.

How It Works

• Multi-platform search: queries web, news, and social sources using keyword groups such as Core, Products, Tech, People, and Variants.
• Semantic deduplication: merges repeated coverage of the same event so reposts do not dominate the report.
• Three-layer sentiment analysis: uses local SnowNLP classification first; confident positives and negatives are handled locally, while uncertain items are escalated to an LLM for precision.
• Heat and alerts: generates heat scores, four-level alerts, and trend data, persisting history in files such as history.json.
• Report delivery: produces HTML and text reports and can send them via AgentMail, which fits scheduled automation.

Boundaries And Notes

The skill depends on Chinese-language content, a local snownlp installation, and keyword quality. Recall and deduplication may weaken when a target is rare, terminology changes quickly, or platforms impose strong anti-scraping constraints. It is better treated as a scheduled monitoring and first-pass screening layer; compliance, legal, or crisis decisions still require human review.