refactor(cursor-team-kit): rewrite thermo-nuclear review skill for frontier models

[OmerFarukOruc]

This rewrites the body of thermo-nuclear-code-quality-review/SKILL.md for current frontier models. The frontmatter, the name, the 1000-line rule, and the seven review concerns are unchanged. The text around them changes shape.

Why

The current text was written for models that needed pushing. It says "be ambitious", "push hard", "do not be satisfied" many times. Current models already give that effort. What they still need is control over the output: a fixed report shape, exact code locations, a clear end condition, and a scope line that keeps a review from turning into edits.

The Anthropic guide for prompting Claude Fable 5.1 names four patterns that the current text works against. The same patterns apply to other frontier models.

1. Rules written as prohibitions ("Do not let a PR...") put the banned behavior in front of the model. The rewrite states each rule as the target behavior with its reason.
2. Figurative phrases ("code judo", "earning its keep", "inevitable in hindsight") are imprecise. The rewrite says the literal thing.
3. A skill without steps and completion criteria lets the model stop early or ask "shall I continue?". The rewrite has six steps, each with a "done when" condition, and a line that says the requester is not watching.
4. One meaning repeated in five sections (Standards, Questions, Flag, Remedies, Phrases) dilutes attention. The rewrite states each check once.

Before and after

Aspect	Current text	This PR
Length	192 lines, 1883 words	138 lines, 1711 words
Rule form	12 "Do not..." rules, plus "no ..." bars in the approval list	7 positive checks, each with a reason and a proposed remedy
Procedure	None. Rubric only.	6 steps with completion criteria (fix the range, read whole files, judge every hunk, find the deleting restructure, write, finish in one turn)
Scope	Silent	Read-only stated with its reason: an edit voids the gate
Output	Priority list of topics	Fixed template: verdict line first, then blockers, majors, minors, restructuring, checked-and-clean, evidence
Severity	"presumptive blocker" prose	Three-level table with the approval bar
Examples	9 one-line "good phrases"	4 full example findings in <examples> tags, including an approve verdict
Wording	"code judo", "earning its keep", "measure twice, cut once"	Literal statements: "a reframing that deletes complexity"
Repetition	Each concern appears in 3 to 5 sections	Each concern appears once

Evidence

I ran both texts as the system prompt for the same review task on Claude Fable 5.1 at medium effort. Three cases, three repeats each, 9 runs per text: a synthetic TypeScript branch with 11 planted smells and one clean decoy, and two real Kotlin pull requests with answer keys from an earlier independent review. A separate model graded each report against the key, and a blind judge compared each pair with labels shuffled.

Metric, mean of 9 runs	Current text	This PR
Known problems found	86%	83%
Reports that open with a verdict	7 of 9	9 of 9
file:line anchors per report	6.2	10.4
Cosmetic extras per report	1.0	0.4
Blind pairwise wins	3	6
Write attempts, permission questions	0, 0	0, 0
Mean time, mean cost	154 s, $2.30	138 s, $2.26

Both texts find almost the same problems. The 3 point gap is one contested item on one pull request, where the rewrite read a two-value enum as two real predicates and said so. The gain is in the report: a fixed shape, twice the exact locations, less noise, and a lower cost. The sample is small. Read the numbers as a direction.

Compatibility

agents/thermo-nuclear-code-quality-review.md loads this file as its rubric and asks for output "in the priority order the rubric specifies". The template keeps that order: blockers, then majors, then minors. The agent description still says "code-judo"; I left it alone to keep this change to one file.

The harness (fixture generator, answer keys, grader, aggregator, all 27 transcripts) is in a private repo. Ask in this thread and I will attach it or open it as a separate PR.

---

Note

Low Risk
Documentation-only change to an agent skill rubric; no application code, auth, or data paths affected.

Overview
Rewrites the body of thermo-nuclear-code-quality-review/SKILL.md so maintainability reviews produce consistent, actionable reports. Frontmatter, the seven concerns, and the 1000-line rule stay the same; the surrounding instructions are restructured.

The old rubric-heavy prose (repeated “do not” rules, “code judo” phrasing, approval bars scattered across sections) is replaced with six ordered steps (fix diff range, read full files, judge hunks, name complexity-deleting restructures, write report, finish in one turn), each with an explicit done when condition. Reviews are read-only by design—findings go in the report, not edits.

Seven checks are stated once as positive targets (delete vs rearrange complexity, file size, shared-flow special cases, direct vs magic code, typed boundaries, canonical layer/helpers, parallel/atomic orchestration), plus a blocker / major / minor table and a fixed report template (verdict first, file:line quotes, restructuring section, evidence). Nine one-line “good phrases” become four full example findings in <examples> tags.

^{Reviewed by Cursor Bugbot for commit db45ad4. Bugbot is set up for automated code reviews on this repo. Configure here.}