Skills
CLAUDE.md Mermaid Diagram Context Compression: Hundreds of Tokens Replace Thousands of Prose
LLMs parse Mermaid diagram syntax significantly more efficiently than equivalent architectural prose: a component diagram or sequence flow that would require 3,000+ tokens of description compresses to 200-400 tokens of Mermaid. Embedding system architecture, data flow, and module relationships as Mermaid blocks in CLAUDE.md gives Claude high-fidelity structural context at minimal token cost. This pairs with the keep-CLAUDE.md-short principle — remove every line that doesn't prevent a mistake, then use Mermaid for what remains.
Source
↳ Follow the thread