What problem does it solve? AI instruction files like CLAUDE.md and AGENTS.md accumulate rules written for older, weaker models, and reading them cannot reveal which rules are now dead weight competing for context. This Skill answers the question experimentally: it runs the same real coding task many times with the instruction layer intact and with it stripped, then grades every rule against what actually changed. ## Core Features & Use Cases - Layer Mapping: Inventories every AI instruction artifact across Claude Code, Cursor, Windsurf, Cline, Copilot, and AGENTS.md, classifying each as always-loaded, on-demand, or enforcement, and reports the per-session token cost. - Automated Ablation Runs: Executes control and stripped arms in throwaway git worktrees built from HEAD, never touching the working tree, with configurable run counts, models, and concurrency. - Blind Grading Rubric: Provides a reference rubric for grading each rule per run as followed, violated, or not applicable, blind to which arm produced each diff. - Use Case: A team suspects their 400-line CLAUDE.md is bloated. The Skill maps the layer, runs a real feature task twice per arm, and produces a per-rule table showing which rules are load-bearing, which the model follows anyway, and which are ignored even when loaded. ## Quick Start Ask the agent to ablate the AI layer of this repository and tell you which CLAUDE.md rules can be safely deleted.
