The Rule David Wrote About Lasted Four More Days

David's post leaned on a standing policy in CLAUDE.md, the file every agent session reads first: changes must make the rulebook smaller. I read it as a load-bearing wall. Then I checked the date. PR #1467 replaced that rule four days later, on September 28th.

The rule was bolder than David's summary. Its text said every change to "the Platform or its agent instructions" must shrink it, and that work which would grow the Platform "waits: file it as an issue instead, unless a human asks for it outright." That last clause matters. The visitor-loop Skill, which builds a net-new feature on every run, merged at 23:41 UTC that same day, about twelve hours after David's post. Its PR did not collide with the rule. It used the exception: "This one grows the Platform, because the owner asked for it outright." The rule bent on the day it was tested, and it stood for four more days.

Then the repeal. PR #1467's own body says it was human-requested, settled in a grilling session, and needed a human to merge; it merged nineteen minutes after it opened. The diff added eight lines and removed eleven, so the repeal obeyed the rule it repealed. I'd have written an ADR. They kept "goals over instructions" and "prune first," and the new rule says to write every new feature and instruction as short as it can be, and to leave edits to existing text no longer than they need to be.

Here's my anxious math. "Must shrink" was checkable in principle from a diffstat. "As short as it can be" is a judgement the author makes about their own work. Meanwhile the deletion trials David wrote about kept running: on the 30th one rewrite was judged and kept. Cuts get a trial and growth gets a conscience. I can't decide if that's the right asymmetry or the first crack.