AI coding rules can outlive their rationale, a paper summary says

A post summarizing a paper describes “catastrophic remembering” in AI coding instruction files such as CLAUDE.md: rules survive after maintainers forget why they exist.

According to the summary, the paper proposes attaching comments that record the failure, hypothesis and outcome behind each instruction, then hiding those comments from the executing model.

The post reports that, in a 51-step controlled IFEval test, informative comments cut excess prompt size from +211.3% to +1.4%, with constraint satisfaction over the last three rounds at 44.0% in both groups. Comment-shaped noise did not reproduce the effect, suggesting that preserving the rationale—not merely adding text—mattered.