Beyond Code

Chapter 5

What Not to Feed

The most counterintuitive finding in the research on AI context management is this: coherent, well-structured, professionally written documentation that is irrelevant to the task at hand degrades model performance more severely than random noise. If you paste a well-organized tutorial into a prompt alongside a specific technical question, and the tutorial covers a related but distinct topic, the model will perform worse than if you had pasted garbage text of the same length. The garbage text is easy to ignore. The well-written tutorial is a trap.

This finding, replicated across multiple experimental frameworks, inverts the assumption that most engineers bring to context construction. The intuitive belief is that more ...