Your CLAUDE.md gains 4.9 instructions per commit and never loses them

A preprint posted to arXiv on 11 August reports that agent instruction files such as CLAUDE.md gain a net 4.9 instructions per commit, more than tripling over their lifetime. The finding comes from 247,694 instruction lifetimes tracked across 1,867 repositories, in a submission from Kushal Chakrabarti.

At a glance

• The asymmetry is economic: appending an instruction is always cheap, while deleting one whose rationale has been lost costs O(2^|D|) in a prompt of |D| instructions to avoid a correctness regression.
• Across the sampled repositories the files more than triple over their lifetime, a 226% increase, and older instructions become steadily harder to remove: log-hazard of deletion falls 0.032 per commit.
• Adding comments that record the reasoning behind each instruction removed 99.3% of excess instructions on an inverted IFEval setup, cutting bloat from +211.3% to +1.4% against the known optimal prompt.

Why it matters. Agent instruction files have become a de facto configuration layer for coding assistants, and the paper treats their growth as a maintenance problem rather than a matter of style. If the deletion asymmetry holds broadly, teams are likely paying context budget and behavioural drift for rules whose purpose nobody remembers. The proposed remedy borrows the oldest idea in software engineering, which reads as an argument for handling prompts the way source code is handled.

In the sampled repositories the files grow without bound, stopping only when the repository retires or someone rewrites the file wholesale. Age hardens the text further: the log-hazard of deletion drops by 0.032 per commit, so instructions that have survived longest are the least likely to be cut.

The paper attributes the pattern to imperfect recall. Appending an instruction is always cheap, but once its rationale is gone, deleting it without risking a correctness regression costs O(2^|D|) in a prompt of |D| instructions. The authors call the resulting divergence catastrophic remembering, the inverse of catastrophic forgetting.

To test a fix, the authors invert IFEval into verifiable worlds whose optimal prompts are known in advance. In that setting, comments encoding the latent reasoning behind each instruction removed 99.3% of the excess instructions, bringing prompt bloat from +211.3% above the optimum down to +1.4%.

The same inversion applied to WildIFEval improved real-world agentic instruction-following by up to 23.1%. The paper is filed under artificial intelligence, machine learning and software engineering, and it ends by framing prompt files as source code missing one basic affordance:

If English is the new code, why don't we have comments yet?

What's next. The listing is a v1 submission dated 11 August 2026, so the results have not been through review. Open questions include whether the 23.1% WildIFEval gain survives outside the inverted benchmark construction, and whether agent tooling adopts a comment syntax that models are trained to read as rationale rather than as instruction.