Best-of-N Agent Code Review: Anthropic Open-Sources Its Skills

This is a short talking-head video. Steve Morin films himself on a San Francisco street, speaking to camera while captions track each phrase. There are no code samples, diagrams, or screen recordings. The only on-screen graphic is an opening title card that reads "AI Code Review Best Practices." The value is in the claim, not a demo: Anthropic has open-sourced the code review skill it uses internally, plus several other skills, and that gives you a working reference for how these are built.

The pitch is two-part. First, a specific code review pattern worth copying. Second, a broader point that Anthropic publishing its own skills lets you read production examples written by the people who designed the format.

The Best-of-N Code Review Pattern

Morin calls the pattern "best of N agent code reviewing" and says it is simple to set up. The captions reinforce the steps as he names them: "N AGENT CODE REVIEWING," then "CHECKING OUT BUGS," then "DUPING OF THESE THINGS."

The setup has two stages:

• Ask Claude Code to spin up multiple parallel agents that each hunt for bugs. Run N copies of that agent.
• Set up a separate set of agents to dedupe the findings across all those parallel runs and check for false positives.

The first stage casts a wide net. Running several independent reviewers in parallel surfaces more candidate bugs than a single pass would. The second stage is the cleanup layer. Because parallel agents flag the same issue more than once and produce some false positives, the dedupe-and-filter agents collapse duplicates and discard the noise before anything reaches you. You get breadth from the fan-out and signal from the dedup-and-filter stage.

Anthropic Open Sourced the Skill

Morin's second point is that this is not just a concept you have to rebuild from scratch. Anthropic has open-sourced the code review skill, along with a number of other skills. The captions land on "OTHER SKILLS," "BY THE ORIGINAL," and "OUT HOW THEY WORK" as he makes the case.

His framing is that people are not realizing all the power behind skills. Because Anthropic built the skills format, reading skills the Anthropic team wrote gives you reference implementations from the original creators. Instead of guessing at structure, you read the real code review skill and see how it works. His closing instruction is direct: go check them out.

Why It Matters

The open-sourced code review skill is both a working tool and a blueprint. You can run it as-is for best-of-N review, and you can study its structure to write your own agentic skills. The released skills serve as a reference for the skill format itself, written by the team that designed it.

Key Takeaways

• Best-of-N code review means running N parallel bug-finding agents instead of a single review pass.
• A second layer of agents dedupes overlapping findings and filters out false positives before you see them.
• You can trigger this directly by asking Claude Code to spin up multiple parallel agents and N copies of the review.
• Anthropic open-sourced its code review skill plus several other skills as a working reference.
• These are examples written by the team that created the skills format, so they show the intended structure and conventions.
• The video is a verbal blueprint, not a walkthrough. There is no code or UI on screen, so the next step is to read Anthropic's released skills directly.