Coding agents read a repository instruction file, and the working assumption is that more detail in it helps. This study generated agent files for all 31 combinations of five content types, used each one on its own to rebuild five real Python libraries from scratch, and ran each library's own test suite against the result. Almost everything compiled — 98.1% for the stronger model, 88.4% for the weaker — and almost nothing worked: 13.0% of the stronger model's runs passed the tests and none of the weaker model's did. The best results came from files holding one, two or three content types rather than all five, and token cost tracked the size of the library instead of the number of categories in the file.