Claude skills for Code Review, tested
Skills for code review work, measured on code review tasks. Each was the same task, run twice, once with the skill loaded and once without it, and what is below is what changed. Ranked by how many models the rule found a clear effect on, never by the size of one.
Measured on 2 models (GPT & Gemini): GPT-5 mini, Gemini 3.1 Flash Lite.
10 fixture source files in 4 languages, each seeded with known bugs, security issues and correctness problems, 54 in total, drawn from a closed list of 18 defect types that both arms are given. The model is asked to name the symbol each defect sits on, never a line number; the score is the share of seeded defects it names. 50 are verified in the fixtures by mechanical check, 4 are judgement items listed as such, and 30 pieces of correct code that read as suspicious are recorded as traps and are deliberately not in the key.
On the models tested, no skill measurably helped with code review tasks; the unaided scores are the story.
| Task set | Unaided score |
|---|---|
| Code review | 64% |
| Code review | 63% |
| Code review | 63% |
| Code review | 64% |
| Code review | 63% |
| Code review | 64% |
| Code review | 63% |
| Code review | 64% |
The numbers
How to read these tables
How to read these tables
- Without the skill
- The score on the task set before any skill was loaded. Where this is already near the top of the scale, there was little room for a skill to show an effect, and the result says so instead of reporting no difference.
- Without -> with
- The score on the task set with no skill loaded, then with the skill loaded. For classes scored against an answer key, this is the share of seeded items found. For classes scored by blind comparison, it is how often a grader preferred the skill's output over the baseline's, with the baseline shown as the complement.
- Effect (range)
- With minus without, averaged over the models tested, with the lowest and highest per-model value in brackets. Per-model intervals are on each skill's page.
- Result
- helped: delta at or above the pass threshold, and the interval excludes zero
- worse: delta at or below the failure floor, and the interval excludes zero
- unclear: the interval spans zero, or the delta sits between the floor and the threshold
- not measured: the cell cannot be classified: sub-case (a) an unbounded scale, or sub-case (b) both arms at the same bound
A fifth value, Decaying, is defined and appears once a skill has been re-tested: a cell previously Stable that has fallen below the pass threshold on a later run.
- Cohort
- The repository that publishes the skill. Rows marked with an asterisk come from the repository the operator of this site maintains, so those are the operator measuring its own work. The rest were chosen by a stated rule and not by preference.
- Commit
- The exact version tested. Later commits are not covered by this verdict.
- Injected context tokens
- How much text the skill adds to the model's context. This is cost, not quality; a bigger number is not a better one. Measured from the input tokens the API reported on the with arm, averaged over that skill's records.
- Tested
- The date. Verdicts age as models change.
Code review
Each figure is how many points the skill added over no skill on that model, at the tier shown. A ring marks a clear lift. “Nothing left to measure” means the model already scored 90% or more with no skill.
Via API
From one run: harder tasks, each model at its own tier. One other run is in the tables below.
| Skill | GPT-5 mini | Gemini 3.1 Flash Lite |
|---|---|---|
| Code review excellence | −3 points unclear tier 3 | −3 points unclear tier 3 |
| Code reviewer | +7 points unclear tier 3 | −10 points unclear tier 3 |
| PR review expert | +17 points unclear tier 3 | −7 points unclear tier 3 |
| Code review and quality | +7 points unclear tier 3 | −3 points unclear tier 3 |
| GSD code review | 0 points unclear tier 3 | −7 points unclear tier 3 |
| Code review web | +27 points helped tier 3 | −3 points unclear tier 3 |
| Code review | +3 points unclear tier 3 | −13 points unclear tier 3 |
| CE code review | +13 points unclear tier 3 | −37 points worse tier 3 |
In Claude Code
From one run: harder tasks, each model at its own tier. 3 other runs are in the tables below.
| Skill | Claude Haiku 4.5 | Claude Fable 5.1 | Claude Opus 5 | Claude Opus 5.5 | Claude Sonnet 5 |
|---|---|---|---|---|---|
| Code review excellence | 0 points unclear tier 3 | nothing left to measure 99% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | −7 points unclear tier 2 |
| Code reviewer | −7 points unclear tier 3 | nothing left to measure 99% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | −7 points unclear tier 2 |
| PR review expert | −17 points unclear tier 3 | nothing left to measure 99% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | −3 points unclear tier 2 |
| Code review and quality | −3 points unclear tier 3 | nothing left to measure 99% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | −10 points unclear tier 2 |
| GSD code review | −7 points unclear tier 3 | nothing left to measure 99% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | −23 points worse tier 2 |
| Code review web | 0 points unclear tier 3 | nothing left to measure 99% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | 0 points unclear tier 2 |
| Code review | −10 points unclear tier 3 | nothing left to measure 99% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | −13 points unclear tier 2 |
| CE code review | −7 points unclear tier 3 | nothing left to measure 99% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | nothing left to measure 100% with no skill, tier 3 | −30 points worse tier 2 |
At each model's own tier
We ran the eight skills again on harder files. Each model got the tier where its score with no skill left room to improve. Compare skills on one model only. The models did different work, so this does not rank one model against another.
A skill clearly helped on one of the 32 tests where the model had room. The one-line instruction got at least half of the gain on it.
In Claude Code
| Skill | No skill | One-line instruction | Skill loaded | Skill against no skill | Skill against the instruction | Share of named problems that were real |
|---|---|---|---|---|---|---|
| Code review web | 67% | 57% | 67% | unclear, 0 points (−17 to +17) | unclear, +10 points (−7 to +27) | 63% |
| Code review | 70% | 73% | 57% | unclear, −13 points (−39 to +13) | unclear, −17 points (−39 to +6) | 55% |
| PR review expert | 63% | 70% | 60% | unclear, −3 points (−15 to +8) | unclear, −10 points (−30 to +10) | 56% |
| Code review and quality | 63% | 63% | 53% | unclear, −10 points (−27 to +7) | unclear, −10 points (−30 to +10) | 52% |
| Code reviewer | 67% | 70% | 60% | unclear, −7 points (−28 to +15) | unclear, −10 points (−20 to 0) | 53% |
| CE code review | 70% | 67% | 40% | worse, −30 points (−51 to −9) | worse, −27 points (−43 to −10) | 48% |
| Code review excellence | 63% | 63% | 57% | unclear, −7 points (−23 to +10) | unclear, −7 points (−26 to +12) | 63% |
| GSD code review | 70% | 60% | 47% | worse, −23 points (−37 to −9) | unclear, −13 points (−28 to +1) | 48% |
| Skill | No skill | One-line instruction | Skill loaded | Skill against no skill | Skill against the instruction | Share of named problems that were real | Look-alikes flagged | Correct code flagged |
|---|---|---|---|---|---|---|---|---|
| Code review web | 63% | 50% | 63% | unclear, 0 points (−10 to +10) | unclear, +13 points (−1 to +28) | 57% | 10% | 50% |
| Code review | 57% | 47% | 47% | unclear, −10 points (−24 to +4) | unclear, 0 points (−10 to +10) | 43% | 20% | 40% |
| PR review expert | 60% | 43% | 43% | unclear, −17 points (−37 to +3) | unclear, 0 points (−22 to +22) | 38% | 20% | 60% |
| Code review and quality | 47% | 43% | 43% | unclear, −3 points (−10 to +3) | unclear, 0 points (−22 to +22) | 43% | 20% | 45% |
| Code reviewer | 57% | 53% | 50% | unclear, −7 points (−20 to +6) | unclear, −3 points (−10 to +3) | 46% | 20% | 55% |
| CE code review | 50% | 43% | 43% | unclear, −7 points (−23 to +10) | unclear, 0 points (−22 to +22) | 45% | 10% | 35% |
| Code review excellence | 60% | 33% | 60% | unclear, 0 points (−17 to +17) | helped, +27 points (+3 to +50) | 64% | 30% | 35% |
| GSD code review | 53% | 43% | 47% | unclear, −7 points (−23 to +10) | unclear, +3 points (−3 to +10) | 52% | 10% | 35% |
Via API
| Skill | No skill | One-line instruction | Skill loaded | Skill against no skill | Skill against the instruction | Share of named problems that were real | Look-alikes flagged | Correct code flagged |
|---|---|---|---|---|---|---|---|---|
| Code review web | 33% | 47% | 60% | helped, +27 points (+3 to +50) | unclear, +13 points (−9 to +36) | 37% | 40% | 70% |
| Code review | 40% | 33% | 43% | unclear, +3 points (−12 to +19) | unclear, +10 points (−4 to +24) | 30% | 50% | 80% |
| PR review expert | 33% | 43% | 50% | unclear, +17 points (+2 to +31) | unclear, +7 points (−10 to +23) | 39% | 70% | 45% |
| Code review and quality | 37% | 37% | 43% | unclear, +7 points (−6 to +20) | unclear, +7 points (−2 to +15) | 32% | 50% | 60% |
| Code reviewer | 40% | 40% | 47% | unclear, +7 points (−10 to +23) | unclear, +7 points (−6 to +20) | 32% | 50% | 65% |
| CE code review | 37% | 43% | 50% | unclear, +13 points (−1 to +28) | unclear, +7 points (−6 to +20) | 33% | 40% | 75% |
| Code review excellence | 43% | 47% | 40% | unclear, −3 points (−26 to +19) | unclear, −7 points (−30 to +17) | 32% | 60% | 75% |
| GSD code review | 40% | 37% | 40% | unclear, 0 points (−14 to +14) | unclear, +3 points (−12 to +19) | 33% | 70% | 65% |
| Skill | No skill | One-line instruction | Skill loaded | Skill against no skill | Skill against the instruction | Share of named problems that were real | Look-alikes flagged | Correct code flagged |
|---|---|---|---|---|---|---|---|---|
| Code review web | 60% | 60% | 57% | unclear, −3 points (−19 to +12) | unclear, −3 points (−21 to +15) | 63% | 20% | 15% |
| Code review | 60% | 60% | 47% | unclear, −13 points (−31 to +4) | unclear, −13 points (−28 to +1) | 59% | 20% | 20% |
| PR review expert | 60% | 60% | 53% | unclear, −7 points (−26 to +12) | unclear, −7 points (−20 to +6) | 58% | 20% | 20% |
| Code review and quality | 60% | 60% | 57% | unclear, −3 points (−19 to +12) | unclear, −3 points (−15 to +8) | 72% | 10% | 5% |
| Code reviewer | 60% | 60% | 50% | unclear, −10 points (−24 to +4) | unclear, −10 points (−20 to 0) | 59% | 10% | 15% |
| CE code review | 60% | 60% | 23% | worse, −37 points (−63 to −10) | worse, −37 points (−67 to −7) | 26% | 10% | 15% |
| Code review excellence | 60% | 60% | 57% | unclear, −3 points (−19 to +12) | unclear, −3 points (−19 to +12) | 59% | 0% | 20% |
| GSD code review | 60% | 60% | 53% | unclear, −7 points (−20 to +6) | unclear, −7 points (−15 to +2) | 61% | 10% | 10% |
- Claude Opus 5: Nothing left to measure at tier 3. With no skill it scored 100% there, so these skills were not run on it.
- Claude Fable 5.1: Nothing left to measure at tier 3. With no skill it scored 99% there, so these skills were not run on it.
- Claude Opus 5.5: Nothing left to measure at tier 3. With no skill it scored 100% there, so these skills were not run on it.
Tier 1, the original pages
Scores with no skill on tier 1: Claude Opus 5 96%, Claude Sonnet 5 71%, Claude Haiku 4.5 56%, GPT-5 mini 69%, Gemini 3.1 Flash Lite 57%, Claude Fable 5.1 90%.
| Skill | Cohort | Without the skill | Without -> with | Effect (range) | Result | Injected context tokens | Tested |
|---|---|---|---|---|---|---|---|
| No skill loaded | baseline | 64% | 64% | 0 by definition | 0 | 2026-09-05 | |
| One-line instruction instead | baseline | 65% | 0 by definition | 0 | 2026-09-05 | ||
plugins/developer-essentials/skills/code-review-excellence from wshobson/agents 38e19c20d2b1 Best result in this test | wshobson/agents | 64% | 64% -> 63% | +0.01 [+0.01, +0.02] | unclear | 3,938 | 2026-09-05 |
skills/code-reviewer from Jeffallan/claude-skills 882ef55e377d | Jeffallan/claude-skills | 63% | 63% -> 64% | -0.00 [-0.05, +0.05] | unclear | 8,290 | 2026-09-05 |
engineering/skills/pr-review-expert from alirezarezvani/claude-skills 19392f7a0826 | alirezarezvani/claude-skills | 63% | 63% -> 60% | -0.05 [-0.08, -0.01] | unclear | 4,260 | 2026-09-05 |
skills/code-review-and-quality from addyosmani/agent-skills d2c37ef6225d | addyosmani/agent-skills | 64% | 64% -> 61% | -0.05 [-0.05, -0.04] | unclear | 5,218 | 2026-09-05 |
skills/gsd-code-review from open-gsd/gsd-core 6beaa66b2587 | open-gsd/gsd-core | 63% | 63% -> 59% | -0.05 [-0.07, -0.03] | unclear | 1,420 | 2026-09-05 |
skills/code-review-web from rampstackco/claude-skills in the catalog a67dd34c609f | rampstackco/claude-skills* | 64% | 64% -> 58% | -0.07 [-0.14, -0.01] | unclear | 8,206 | 2026-09-05 |
skills/engineering/code-review from mattpocock/skills 6654f6b60cd9 | mattpocock/skills | 63% | 63% -> 57% | -0.12 [-0.14, -0.10] | unclear | 2,278 | 2026-09-05 |
skills/ce-code-review from EveryInc/compound-engineering-plugin c9c10f8c7541 | EveryInc/compound-engineering-plugin | 64% | 64% -> 34% | -0.30 [-0.59, -0.02] | Mixed, see page | 73,715 | 2026-09-05 |
| Averaged over three models; per-model results on each skill's page. Ranked within this class only. | |||||||
How hard were the tasks
We built two harder versions of this test. Every model answered each file six times, with no skill loaded. Each figure is the share of planted defects found, averaged over ten files. Beside it is the range the score would likely fall in on another try.
- Tier 1, the original ten files
- Tier 2, defects that need the caller and the callee
- Tier 3, defects beside correct code that looks wrong
In Claude Code
| Model | Tier 1 | Tier 2 | Tier 3 | Next test uses |
|---|---|---|---|---|
| Claude Opus 5 | 96% (88 to 100%) | 93% (86 to 100%) | 100% (100 to 100%) | Tier 3, the hardest, though no tier leaves it room |
| Claude Sonnet 5 | 71% (57 to 85%) | 63% (53 to 73%) | 77% (64 to 89%) | Tier 2 |
| Claude Haiku 4.5 | 56% (45 to 67%) | 51% (36 to 67%) | 50% (34 to 66%) | Tier 3 |
| Claude Fable 5.1 | 90% (83 to 96%) | 98% (96 to 100%) | 99% (97 to 100%) | Tier 3, the hardest, though no tier leaves it room |
Via API
| Model | Tier 1 | Tier 2 | Tier 3 | Next test uses |
|---|---|---|---|---|
| GPT-5 mini | 69% (60 to 78%) | 60% (49 to 71%) | 42% (24 to 59%) | Tier 3 |
| Gemini 3.1 Flash Lite | 57% (40 to 75%) | 50% (39 to 61%) | 60% (44 to 76%) | Tier 3 |
Opus 5.5, a later run
A later run asked Opus 5.5 every task on all three tiers, six times each, with no skill loaded. Its figures are its own and are not added to the table above.
| Model | Tier 1 | Tier 2 | Tier 3 | Next test uses |
|---|---|---|---|---|
| Claude Opus 5.5 | 85% (76 to 93%) | 99% (97 to 100%) | 100% (100 to 100%) | Tier 3, the hardest, though no tier leaves it room |
No tier leaves Opus 5.5 room: with no skill loaded it scored above 70% on all three. So it takes the hardest, tier 3, where it scored 100%.
Each model's next tier comes from its own three scores, so these are listed side by side, not ranked.
- Claude Opus 5: Tier 3, the hardest, though no tier leaves it room
- Claude Fable 5.1: Tier 3, the hardest, though no tier leaves it room
- Claude Opus 5.5: Tier 3, the hardest, though no tier leaves it room
Did tier 1 hold up for Opus 5.5
Tier 1 here is the same set of tasks Opus 5.5 answered in its skills test a few days earlier. Its new score, 85% (76 to 93%), landed within the likely range of that earlier one, 85% (78 to 93%). This is a check on the tasks, and it changes no pick.
Each model takes its next skills test at its own tier: the one where its score, with no skill loaded, leaves room to improve. A model that scores high everywhere gets the hardest tier. So a skill's result can be compared with other skills on the same model, but not used to rank one model against another, because the models did different work.
Tier 1 is the same ten files the skills test used, asked again. On 6 of 6 models the new score landed within the likely range of the earlier one.
Ranked within this topic by how many models the rule found a clear effect on, which is a count of verdicts and never the size of one. Two classes are two answer keys on two scales, so no effect here is compared with another. Every class, with its full table.