At the beginning of each session, an LLM agent loads a fixed context file, for example $\texttt{AGENTS.md}$. Every token loaded from the file is charged again in later rounds, and as the file grows its size can degrade performance. In practice, human or automated curators usually grow these files by appending.
We formulate context curation as a capacitated assortment problem. Instructions consume a finite attention capacity; adding an instruction never raises the compliance of the others, while retained instructions incur a per‑session setup cost.
We prove an upper bound on the optimal file size regardless of how many candidate instructions are available, and show that appending every instruction with positive standalone value can be arbitrarily worse in net value than selecting an optimal subset. A token budget also limits loss when the token price is underestimated.
We then examine how past sessions can be leveraged to guide addition or removal of instructions. Feedback is inherently censored: the benefits and harms of loaded instructions are observable, whereas missing instructions generate feedback only when their absence causes harm. In this setting, deleting instructions ignored by agents inevitably removes helpful ones. We characterize how much evidence should be collected before adding an instruction and bound regret when human reviewers can inspect only a limited number of edits per period.
Empirical experiments further demonstrate that irrelevant rules drawn from real context files reduce language‑model compliance.
Review: The study offers theoretical limits on context file size and quality, highlights the risks of naïvely appending instructions, and provides practical guidelines for effective curation.