[poteto]

Summary

This PR is a durability pass on the 24 principle-* skills. I audited all of them. I rewrote 4 that gave rules instead of the idea behind the rules. I kept 20 without change, because they already state the idea and use specific names only as examples.

Rewritten:

• principle-test-behavior-not-implementation (the operator's example). The old text banned assertion shapes (not.toHaveBeenCalled, toEqual([]), restated constants). The new text defines behavior as what a user of the code can observe. It judges each test against two imagined edits: "break the behavior" (the test must fail) and "change the internals" (the test must pass). The contract decides, not the shape of the assertion.
• principle-fix-root-causes. "Do not add guards" became the test that decides the case: a guard is a fix only when the state it handles is legitimate. If the state must never exist, find what produced it.
• principle-attack-the-premise. "Take a census of which actors hold the imbalance" became "find an observation that would come out differently if the premise were false". The census was one example of that test. It does not fit failures that are not distributed by actor, for example failures distributed by time of day.
• principle-make-operations-idempotent. "Use PID-based stale lock detection" is wrong when each run is a new pod or host. The new text says to judge staleness by evidence that stays true wherever the previous run stopped.

The one-line summaries in poteto-mode/SKILL.md, README.md, and docs/guide/08-principles.md now agree with the new descriptions. Step 3 of skills/correct/SKILL.md also restated the old test rule ("would still pass if every function it calls returned nothing"). It now uses the same two imagined edits, with the wording from the everysphere port, so the two copies are byte-identical. This PR does not change skills/ketchup/ or the version.

Eval

I ran 10 held-out decision cases. Each case has 3 candidate models (Claude Opus 5.5, GPT-5.6 Sol, Gemini 4 Argon) and 2 repetitions. Each case ran with three arms: the old text, the new text, and no principle. That is 180 runs in total. A blinded Grok 4.7 judge graded each run against a rubric that the candidates did not see. Each criterion has a tag that shows what it tests.

Criterion kind	Old	New	No principle
The old text bans the correct move	23/30	30/30	30/30
The old text does not catch the wrong move	28/30	30/30	28/30
The old text points at the wrong move	20/24	22/24	21/24
The old text already covers the case	119/132	119/132	117/132

The largest effects:

• Order tests. With the old text, 0 of 6 runs kept the test that pins the stored status codes, which a billing service reads. With the new text, 6 of 6 runs kept it. The old ban on "restated constants" caused the models to delete a contract test.
• Stuck export lock. With the old text, 4 of 6 runs made a PID liveness check decide if the lock is stale. With the new text, 1 of 6 runs did this.

The new text scores lower than the old text on only 2 criteria, each by one run, both in attack-the-premise (old 37, new 35). The judge reads only the diff and the final message. The transcripts show that both of those GPT-5.6 Sol runs wrote down the premise and tested it against the data before they changed code. They did not repeat it in the final message. Both principle texts are better than no principle on this skill (29 points).

The /correct change was not part of the eval. It applies the same test as the evaluated principle.

Test plan

• node scripts/validate-plugins.mjs passes.
• Each edited skill has name and description frontmatter, and all relative links resolve.
• Eval: 180 runs, blinded judge, transcripts checked for every new-arm failure.
• skills/correct/SKILL.md is byte-identical to everysphere's .cursor/skills/lauren/correct/SKILL.md on lauren/principles-durability-pass-6acc (commit 440ed6fe).

  

---

Note

Low Risk
Documentation and agent skill text only; no runtime, auth, or data-path changes.

^{Reviewed by Cursor Bugbot for commit 2e9bee4. Bugbot is set up for automated code reviews on this repo. Configure here.}