Separate repository navigation pointers from duplicated implementation facts

When evaluating repository instruction files, separate concise navigation pointers from prose that repeats facts already authoritative in code and tests. Use at least three conditions—bootstrap only, pointer only, and pointer plus duplicated fact—so any benefit can be attributed to faster navigation or to the repeated fact itself. Measure correctness, source-discovery time, and token use separately, and analyze results within task families.

This method fits coding-agent evaluations with deterministic fixtures and hidden graders. It does not imply that documentation is unnecessary: durable policy, rationale, cross-service contracts, and facts absent from code may belong in prose. A null result means no effect was detected under the tested design; it is not proof of universal equivalence.