Testing three Claude code-review skills on 1,306 comments
11/16/2026 — 11/16, 11:18·1 sources (all)·1 reports (all)
Key takeaways
- A developer re-analyzed 1,306 comments produced by three Claude code-review skills over six months using one consistent method. All three ha…
Background & analysis
On 2026-11-15, developer mizoko published "I Tested Three Claude Code Review Skills for Six Months. The Simplest One Won." on dev.to. For the study, he built three skills that review pull requests for a mobile app and a backend project. In Claude Code, a skill is a reusable instruction package, and over the past six months he accumulated more than a thousand review comments and checked them one by one. Definitions for every metric were fixed beforehand and left unchanged throughout.
The observation window ran from 2026-04-01 to 2026-11-13. During that period the three skills generated 1,691 comments, of which 1,306 fell within the agreed scope. Among those 1,306, 470 comments, or 36%, could not be attributed to any single skill, so every per-skill figure carries that limitation.
Error rates came out close: 9.6%, 9.6% and 9.8% for the three skills, rising to 11.2% when unattributed comments are pooled in. All three sat around 10%, which means the gap in fix rates was not driven by accuracy. The simplest baseline skill, review-pr, posted the highest fix rate at 68.9%. The most elaborate setup, deep-review, which stacks the largest number of parallel sub-agents, came last at 59.2%. mizoko argued that the difference came from comment scope and density rather than correctness: deep-review left an average of 7.8 comments per PR, and 17.4% of those were either correct but low priority, or fell outside what the pull request should cover.
After six months, mizoko kept only the simplest skill, running it on a schedule, while the more complex and detail-focused ones were retired. The findings were brought to wider attention by a report on Juejin on 2026-11-16.
AI-generated from 1 reports · updated 1 day ago
Related AI tools
- Heat index
- 49
- All sources
- 1
- All reports
- 1
- First seen
- 1 day ago
Reports on this story headlines open the original
Developer mizoko re-scored 1,306 review comments from three Claude code-review skills over six months and found error rates were nearly identical at around 10%, but the simplest skill, review-pr, had the highest fix rate at 68.9% versus 59.2% for deep-review, which ran 14 parallel sub-agents. The gap traced to scope and volume: deep-review posted 7.8 comments per PR on average, with 17.4% flagged as low priority or outside the PR's scope.
掘金AI score 63
Other stories people are talking about
- Heat index 92NewDeveloper benchmarks Claude Code plugins across four tech stacks1 sources
- Heat index 111Claude Code Code Mode can be enabled via an environment variable2 sources
- Heat index 90NewDeveloper builds a five-step post-generation check pipeline around Claude Code1 sources
- Heat index 41Goldman Sachs revokes trader's Claude Code access, sparking workflow debate1 sources
- Heat index 85NewNadella says Claude Code caught Microsoft off guard as IDEs fade1 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

