A new AI skill called token-saviour cuts codebase-reading costs by ~66% by routing tasks to the right tool instead of reflexively reading whole files.

What happened: A methodology was benchmarked on a ~30-file Python app showing that swapping whole-file reads for semantic retrieval reduced total tokens by ~66%. The approach uses four specialized tools—serena or graphify for code navigation, rtk for command output, and caveman for trimming verbose replies—each handling a different layer of token cost.

Why it matters: When an AI agent works in a codebase, input dominates the token bill (~88% in the benchmark), and most of that waste comes from over-reading. Routing the right query to the right tool (e.g., graphify for tracing a 4-layer call path at 65 tokens vs. 1,633 for reading files) dramatically reduces cost without losing understanding, freeing tokens for the actual task.

What to watch: The recommended stack—(serena or graphify) + rtk + caveman—achieved a combined −70% token reduction. The tools are independent: each owns one layer, so they stack additively. The approach trades reflexive reads for semantic queries, making it most valuable for large codebases where reading overhead is highest.