Probabilistic benchmarking of .claude/ instruction sets for Claude Code token efficiency
Browse cluster: Cluster 140 →This project is a benchmarking harness that measures token efficiency across six different `.claude/` instruction set configurations for Claude Code. It tests each configuration against three coding challenges (SQLite window functions, a Hono WebSocket counter, and a Python CSV reporter), measuring both token consumption and task completion rates. The project is built in Bash and Python, using git worktrees for isolated test environments and automated verification scripts, and was built by the repository author using Claude Code itself. The work emerged from a Hacker News debate about whether aggressive instruction-set rules actually reduce token usage, leading the author to empirically test six configurations including the actual 61-line CLAUDE.md file from drona23/claude-token-efficient alongside simpler alternatives.