[jasperboerhof]

Why

We measured the patching problem before settling on a change. Four auditors replayed 13 long review chains in kendo, emmie, lokalekeuze and crit: 144 crit rounds, and 230 findings filed in round 2 or later. They tested a set of stop rules against them: a "same bug class at a second site" stop, a blame check, a per-file round counter, and a sweep for twins.

The stop rules would have saved about 7–30 of the 144 rounds, and they fired falsely about half the time. Detection was not the gap. In every repo, the old process already stopped at about the same rounds. What decided the outcome was the fix that came next:

• About half the later findings (111 of 230) sat inside the previous fix. That fix had added a flag, a guard, a lock or a mode, and the new state had holes of its own.
• The chains ended when a fix deleted something: one owner for a piece of state instead of two, a field taken out of a payload, a construct replaced, or one decision recorded for a whole class.
• Design calls that chose to add state kept the chains going.

So this PR drops the stop rules and changes only how a fix is made.

What

skills/shepard/SKILL.md, steps 4 and 5 only:

1. "Make the minimal fix" becomes "Leave the code simpler, not the diff smaller." When a fix adds a flag, guard, lock or mode, look for the version that does not need it. Unrelated refactoring still stays out of the diff.
2. New: "Think the fix through before you push it." Three questions about the fix itself: what new states or values it adds, what happens when the call fails, overlaps, retries or runs twice, and who else reads what it changed. Pin the failure path with a test. A rushed fix costs a whole review round.
3. Step 5: between fix shapes, recommend the one that leaves less state.

The branch's first commit added the stop rules, a Class: commit line and a Hazards section in the template. The second commit reverts all of that. The net diff against main is only the three points above.

Not in this PR

• PR size. In the same data it is the strongest predictor of long chains: PRs under 200 lines averaged 1.4 crit rounds, PRs over 2,500 lines averaged 6. That belongs in the build and plan steps, not in /shepard.
• Late discovery. Many extra rounds came from the reviewer finding old code late. No author-side rule fixes that.
• Drift. The catalog and personal copies still differ elsewhere, in both directions, mostly in the scripts. A resync is its own PR.
• Mirror. lokalekeuze carries a copy of this skill in .claude/skills/shepard. It needs the same change after this merges.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AiJB5Aakjd3rGcm6jdNf3y