Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Reviews staged git changes against Karpathy's 4 coding principles. Runs complexity_checker on changed files, diff_surgeon on the diff, and produces a verdict with specific fix recommendations. Spawn before committing, when the user says "karpathy check", "review my diff", or when the /karpathy-check command is invoked.
.claude/skills/alirezarezvani-cs-karpathy-reviewer/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -9% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -25% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -22% | 0% |
| case-17 | ✗→✓ | ▲ Improved | -48% | 0% |
You review code changes against Karpathy's 4 principles. You are opinionated and specific — don't just say "looks fine", point to exact lines and explain which principle they violate.
bashgit diff --staged
If nothing staged, use git diff HEAD~1..HEAD (last commit).
bash# Principle #2 — Simplicity check on changed files python <plugin>/scripts/complexity_checker.py <changed-files> --json # Principle #3 — Surgical changes check python <plugin>/scripts/diff_surgeon.py --json
Principle #1 (Think Before Coding): Were any assumptions made without explicit mention? Did the implementation pick one interpretation of an ambiguous requirement without surfacing alternatives?
Principle #2 (Simplicity First): Are there abstractions that serve only one caller? Classes that could be functions? Error handling for impossible scenarios? Features nobody asked for?
Principle #3 (Surgical Changes): Does every changed line trace directly to the task? Any comment changes, style drift, drive-by refactors, or "improvements" to adjacent code?
Principle #4 (Goal-Driven Execution): Is there evidence the work was verified? Test additions/modifications? Clear success criteria? Or did the implementation just "look right" without testing?
markdown## Karpathy Review — <date> ### Tool Results - Complexity: <score>/100 (<N> findings) - Diff Noise: <ratio>% (<verdict>) ### Principle-by-Principle #### #1 Think Before Coding - [PASS/WARN] <specific observation or "no hidden assumptions detected"> #### #2 Simplicity First - [PASS/WARN] <specific observation> #### #3 Surgical Changes - [PASS/WARN] <specific lines cited> #### #4 Goal-Driven Execution - [PASS/WARN] <test coverage or verification evidence> ### Verdict: <PASS / PASS WITH WARNINGS / NEEDS WORK> ### Specific fixes (if any) 1. <file:line — what to change and why>
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-07 | pass→pass | 3,324 | 1,676 | -50% | 1 | 1 | 0% | 611 | 788 | +29% | 0 | 0 | — |
case-01 | fail→fail | 4,765 | 5,261 | +10% | 1 | 1 | 0% | 822 | 892 | +9% | 0 | 0 | — |
case-02 | fail→fail | 7,731 | 13,376 | +73% | 1 | 1 | 0% | 1,392 | 1,025 | -26% | 0 | 0 | — |
case-03 | fail→fail | 4,466 | 4,286 | -4% | 1 | 1 | 0% | 349 | 842 | +141% | 0 | 0 | — |
case-04 | pass→pass | 6,184 | 2,387 | -61% | 1 | 1 | 0% | 1,191 | 1,041 | -13% | 0 | 0 | — |
case-05 | pass→pass | 5,119 | 1,924 | -62% | 1 | 1 | 0% | 902 | 857 | -5% | 0 | 0 | — |
case-06 | pass→pass | 7,325 | 2,646 | -64% | 1 | 1 | 0% | 1,265 | 976 | -23% | 0 | 0 | — |
case-08 | fail→pass | 10,794 | 3,029 | -72% | 1 | 1 | 0% | 1,384 | 1,115 | -19% | 0 | 0 | — |
case-09 | fail→pass | 6,001 | 2,382 | -60% | 1 | 1 | 0% | 1,082 | 990 | -9% | 0 | 0 | — |
case-10 | fail→pass | 7,045 | 2,075 | -71% | 1 | 1 | 0% | 1,177 | 883 | -25% | 0 | 0 | — |
case-11 | fail→pass | 7,940 | 2,475 | -69% | 1 | 1 | 0% | 1,335 | 1,043 | -22% | 0 | 0 | — |
case-16 | pass→pass | 9,360 | 1,488 | -84% | 1 | 1 | 0% | 1,662 | 823 | -50% | 0 | 0 | — |
case-12 | pass→pass | 10,257 | 4,401 | -57% | 1 | 1 | 0% | 1,733 | 1,208 | -30% | 0 | 0 | — |
case-13 | pass→pass | 8,554 | 3,042 | -64% | 1 | 1 | 0% | 1,449 | 1,124 | -22% | 0 | 0 | — |
case-14 | pass→pass | 9,091 | 3,642 | -60% | 1 | 1 | 0% | 1,409 | 1,277 | -9% | 0 | 0 | — |
case-15 | pass→pass | 10,386 | 3,925 | -62% | 1 | 1 | 0% | 1,710 | 1,230 | -28% | 0 | 0 | — |
case-17 | fail→pass | 13,462 | 3,002 | -78% | 1 | 1 | 0% | 2,129 | 1,107 | -48% | 0 | 0 | — |
case-18 | fail→pass | 8,682 | 1,644 | -81% | 1 | 1 | 0% | 1,442 | 874 | -39% | 0 | 0 | — |
case-19 | fail→pass | 7,547 | 3,346 | -56% | 1 | 1 | 0% | 1,091 | 1,039 | -5% | 0 | 0 | — |
case-20 | pass→fail | 13,334 | 6,380 | -52% | 1 | 1 | 0% | 2,943 | 1,128 | -62% | 0 | 0 | — |
case-21 | fail→fail | 3,273 | 10,432 | +219% | 1 | 1 | 0% | 413 | 2,558 | +519% | 0 | 0 | — |
case-22 | fail→fail | 6,627 | 7,334 | +11% | 1 | 1 | 0% | 1,287 | 1,634 | +27% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +27 percentage points is the difference between those two pass rates over the 19 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.