One skill, for the work that comes after the diff is understood: sweeping the codebase for the rest of the family the bug in front of you belongs to.
Generic code review is a solved problem and you should not install a skill for it. Claude Code ships a review command, it is good, and a persona file telling a model to act as a senior engineer adds nothing on top of it.
What is not solved is repetition. A bug reaches review because somebody held a wrong belief or copied a neighbouring file. Neither of those causes produces exactly one bug. But the reviewer is looking at a diff, the diff contains one instance, and so one instance is what gets fixed. The rest of the family ships. Three weeks later the same vulnerability is reported against a different endpoint.
We wrote two skills for this category and tested both twice. Only one survived, and it is worth being specific about the one that did not, because it is the more interesting result.
The one that failed was about grounding identifiers. The theory was that every column name, route and environment variable in a change is an unverified claim about a file somebody else owns, and that nothing checks those claims. The theory is right. The skill still did not help: on a real codebase carrying real drift, including a field name that exists on the API types and does not exist on the database model, the assistant scored 14 of 18 without the skill and 14 with it. It already checks, when the schema is in front of it.
That result matters more than it looks, because the first time we tested that skill our control arm was contaminated. It had quietly inherited a standing instruction to verify database field names, which is the skill's entire thesis, sitting inside the arm that was supposed to be working without it. We found the contamination, fixed it, re-ran on a genuinely clean control, and the skill still did nothing. The old number was untrustworthy. The old conclusion was right.
The one that worked was about repetition, and it only worked once the test got harder: a real repository instead of a fixture, and the prompt a person actually types instead of a well-formed question. On that, the floor moved a long way. It also costs something real, which its page states plainly.
Each skill was given a realistic task, then the identical task was run again with the skill removed. The two outputs were compared against a fixture with a known set of planted defects, so the difference is countable rather than a matter of opinion. If a skill did not make the output materially better, it is not on this page.
Each skill page carries the specific result: what the run without the skill found, what the run with it found, and where it still falls short. Back to all skills.