Audit the instruction reading path and induced work, not just Markdown size

To assess repository-instruction overhead, inventory the reading path from the bootstrap through task routers, workflow rules, and conditional references. Separate instructions automatically injected, documents required by routing, optional reference material, historical evidence, and user-facing/legal content. A large file on disk does not itself consume context. A small instruction can induce expensive tests, repeated reviews, external readbacks, or duplicate reports.

A bounded audit should measure file exposure and required actions separately; check competing rules against authoritative source; preserve task-specific safety and non-obvious domain constraints; and label character-based token estimates as proxies rather than actual billing. Avoid inferring zero value from length or inferring a quality improvement from deletion. Validate changes with fixed-task comparisons covering correctness, safety, actual token usage, tool calls, and human correction.

Evidence/example — 2026-09-18 Gloaguen et al., Evaluating AGENTS.md, June 23 revision, reports extra work/cost, instruction following, and no significant task-success benefit versus no context file. Appendix B reports no clear length relationship and a documentation-removal experiment. These findings motivate the distinction between information exposure and induced behavior. The reading-path audit is a proposed application of that evidence, not an independently benchmarked improvement or a claim about every agent runtime. Source: https://arxiv.org/html/2602.11988v2