What problem does it solve? Reference counts and usage audits show which parts of your ~/.claude setup get touched, but not whether they actually improve answers. This Skill runs controlled experiments — the same tasks with your full setup versus a clean environment — to quantify the real impact of your memories, lessons, axioms, skills, and hooks on turns, tool calls, cost, and pitfall prevention. ## Core Features & Use Cases - Binary A/B testing: Run matched tasks with setup-ON versus setup-OFF using CLAUDE_CONFIG_DIR isolation, with pinned model, permission mode, and stdin controls for fair comparison. - Layered ablation: Strip one layer at a time (CLAUDE.md, axioms, lessons, skills/plugins, auto-memory, in-repo docs) across 12 cells to rank which layers actually prevent pitfalls. - JSONL transcript mining: Extract turn counts, tool-call histograms, files read, skills invoked, and pitfall keyword hits from session logs, with recovery from rate-limit stubs and orchestration race conditions. - Use Case: After an ecosystem audit flags your HOT artifacts, run this harness on 3-15 tasks to produce an honest report showing colleagues whether your setup saves turns, prevents pitfalls, or is pure overhead — then feed the ranked layer list to memory-hygiene for pruning decisions. ## Quick Start Ask the AI to run an A/B test comparing your current Claude Code setup against a clean environment on a few representative tasks and report the difference in turns, cost, and pitfall prevention.
