How LLMs Read Docs (www.aiwiki.dev)

The note outlines a practical problem: LLM-based coding agents currently bloat their contexts by ingesting huge, messy documentation files (examples >380k tokens) during pretraining or retrieval, much of which is irrelevant, outdated, or even unusable (e.g., React components, large images, or scraped HTML). That wastes tokens and compute (Claude Code’s ~200k token context is cited), forces repeated web scraping across many agents, and produces unreliable results because pages may not exist or are poorly formatted. Provider-specific tools (MCPs) and ad-hoc scrapers add further fragmentation and inconsistency. The proposed solution is a single, structured “read” tool / unified docs graph that returns the exact snippet an agent needs (e.g., read(shadcn/directory/…/cards) yields a small, ready-to-use markdown block with tags, usage, and install commands). Technical implications: combine canonical documentation into a post-processed, chunked knowledge source with deterministic routes plus vectorized search fallbacks for unknown paths; validate existence and provide fallback scenarios; and standardize read schemas to avoid per-provider tool bloat. Though building this consolidated index is compute- and labor-intensive, much of the preprocessing (formatting, chunking, deduplication) can be automated with AI, reducing redundant scraping, cutting context size, improving reliability, and lowering LLM inference costs for coding agents.