Web search helps but has drawbacks. For individual developers: each search adds latency and compute costs (API calls, processing, synthesis). The training dataset aggregates all that searchable information once, structures it for AI ingestion, then eliminates redundant processing. Instead of search-wait-paste cycles with inconsistent fragments, you get comprehensive framework knowledge instantly. This captures deeper insights (like constrained decoding architecture and ToolOutput patterns) that aren't easily found through search.

Front-loading the knowledge processing means every subsequent interaction is faster and lighter than triggering new searches.

This touches on an interesting concept - knowledge arbitrage - this information has a shelf life, once the SOTA LLMs know about iOS 26 my files turn into the Beta Max of training data.

Here's how i did it:

https://rileygersh.medium.com/how-i-gave-claude-gemini-knowl...