last30days-skill: The High-Stakes Architecture of Real-Time Research
How a modular signal engine triangulates truth across prediction markets, social sentiment, and the open web to kill the LLM knowledge cutoff.
- Prediction market data from Polymarket grounds social media hype with financial stakes.
- An intent-aware router dynamically selects specific platforms like YouTube or Hacker News based on the query type.
- Logarithmic scoring prevents viral engagement metrics from overwhelming smaller but highly relevant data points.
- The engine bypasses traditional search APIs to extract raw sentiment from unfiltered social transcripts and comments.
The Financialization of Fact
The knowledge cutoff is the original sin of Large Language Models. While most developers attempt to fix this with Retrieval-Augmented Generation (RAG) or basic Google Search tools, last30days-skill treats the internet not as a library, but as a high-frequency trading floor.
By treating Polymarket as a primary signal alongside Reddit and YouTube, this project codifies the search habits of an investigative researcher into a deterministic Python engine. It is a deep research tool that values skin in the game over SEO-optimized text.
AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary
The Intent-Aware Router
Instead of hitting every API for every query, the system uses a heuristic intent classifier. The query_type.py module programmatically decides that a product search requires YouTube transcripts and Reddit threads, while a concept search relies on blog posts and Hacker News.
Scoring the Noise
Raw API results are invariably noisy. The score.py module implements a multi-factor scoring algorithm to rank items by relevance, recency, and engagement. Crucially, it uses logarithmic scaling for engagement metrics.
This mathematical approach ensures a viral tweet with one million likes doesn't completely drown out a highly relevant Reddit comment with one hundred upvotes. Truth, in this engine, is a weighted average of convergence across platforms.
| Feature | last30days-skill | Standard RAG |
|---|---|---|
| Primary Signal | Social/Financial (Reddit, Polymarket) | SEO/Web (Google, Bing) |
| Temporal Window | Hard 30-day decay curve | General or arbitrary |
| Truth Mechanism | Cross-platform triangulation | LLM summarization of top links |
Beyond the Web: The Social Scrapers
To access unfiltered discourse, the project bypasses official, highly restricted APIs. It uses tools like yt-dlp for YouTube transcripts and ScrapeCreators for Reddit and TikTok, extracting raw human sentiment before it is packaged by search engine algorithms.
Finally, a hybrid deduplication system uses token Jaccard similarity to prevent the echo chamber effect, ensuring the final synthesized report provides a diverse, high-signal briefing of the modern internet's true pulse.