Your CLAUDE.md grows because nobody wrote down why each rule is there

Chakrabarti tracked individual instructions, not file sizes, across 1,867 GitHub repositories and 247,694 instruction lifetimes. At its last tracked version the median agentic context file carries 39 instructions, with a 90th percentile of 131. Instruction count more than triples over a file's own lifetime, +226%, and grows by a net +4.9 instructions per commit across 19,267 commits once mass rewrites are excluded.

Deletions do happen, but almost never one at a time: 77.3% of instruction deaths arrive in a wholesale rewrite or a migration to a sibling file. Aligning files on their first such rewrite, the mean instruction count drops to 59.5% of its pre-rewrite value and is back to 91.5% within ten commits, growing faster afterwards than before, 4.9% per commit against 4.1%.

The mechanism claim rests on the shape of the deletion hazard. If instructions were being removed because they had gone stale, removal would get more likely with age. It gets less likely: a log-hazard slope of -0.032 per commit, 95% CI [-0.047, -0.019], over 28,426 tracked deletions, and the hazard falls further the more authors have touched the file. The paper's framing of the asymmetry is direct: "appending an instruction is always cheap, but once an instruction's rationale is gone, deleting it without risking a correctness regression costs O(2^|D|) in a prompt of |D| instructions."

The proposed fix is a comment channel. Each instruction carries a note recording the failure that prompted it, the hypothesis, and how the hypothesis has fared. The harness strips the comments before the prompt reaches the executor, so they are addressed to the next maintainer only. In constructed test worlds whose optimal prompt is known by design, excess prompt size falls from +211.3% to +1.4% over 51 maintenance rounds at parity constraint satisfaction. On prompts drawn from WildIFEval, comments lift constraint satisfaction from 50.4% to 62.0% over three rounds, 11.6 points. A placebo arm of comment-shaped text with no real reasoning in it lands within noise of the control, which is the ablation that makes the result about the content of the comment rather than its presence.

The satisfaction gain is a rate under an LLM judge; re-scoring all 6,336 verdicts under a second judge put the effect at 7.8 points instead of 11.6, with a confidence interval that includes zero. The size result comes from worlds whose optimal prompts hold two or three instructions, against a median real file's 39.

Why it matters: every practitioner running agents at scale maintains one of these files, and most of them have been told to prune it without being told how to decide what to cut. The difficulty comes from missing information, not laziness, and that information is cheap to record at the moment you add the rule. If the finding holds, the next feature to want from a harness is a comment syntax that the executor never sees.