Measure context and move procedure into a skill

Make the experiment measurable before you run it. A generator buries clause 6.4 in the middle of a long lease next to similar-sounding clauses, a budget script shows what the pack costs, and a small harness runs the question through claude -p in both conditions and grades the answer. The decision that matters most is where each instruction lives: CLAUDE.md holds only standing rules (about 190 words, paid for on every turn), and the read-lease skill holds the procedure (synonyms to search, a window to read, exceptions to check), paid for only when it's used.