Review skill links

The scripts catch what can be matched by text. This review catches what needs reading: a skill nobody routes to in the one place it matters, two skills giving opposite advice, the same rule in two places. Follow the "Skill boundary discipline" section of CLAUDE.md throughout.

1. Run the checks

scripts/check-skill-refs.sh
scripts/check-skill-wiring.sh
scripts/check-version.sh
scripts/check-frontmatter.sh
scripts/check-skill-callers.sh
scripts/check-evals.sh

Fix or report every failure before reading further. A failure here is a finding, not background.

2. Pick the scope

Find what changed since the main branch:

git diff --name-only "$(git merge-base HEAD origin/main)" -- plugins/

If the user names skills, start there. Otherwise start with the changed and new skills. Either way, read each one's frontmatter description and section headings, then find its neighbors: skills that cover nearby ground (umbraco-backoffice-conventions next to umbraco-backoffice-ux, umbraco-performance next to umbraco-extensibility, and so on).

3. Read for these problems

For each skill in scope and each neighbor:

• Routing gaps. Every list that routes work to skills should name the skill when its territory comes up: agents/builder.md, agents/reviewer.md, feature-design's skill list, quick-fix step 3, feature-build's builder brief, investigate-bug's suspects, and the README tables. Check each list against the skill's description, not just whether the name appears somewhere.
• Conflicting advice. Two skills telling the reader to do opposite things in the same situation. Read the rules on shared topics (caching, background work, handlers, file access, versioning) side by side.
• Duplicated rules. The same rule stated in two skills. Keep it in the skill whose description it fits; replace the other copy with a one-line pointer.
• One-way pointers. If skill A says "for X, see B", check whether B's matching section needs "for Y, see A". Not every pointer needs a return one; add it when a reader of B would otherwise miss A.
• Stale facts after a move. A rule that moved to a new skill but is still described in its old place, or a catalogue entry that now contradicts the new skill.
• Caller checks. A skill that describes its caller or acts differently depending on which skill started it, in wording the script didn't catch. Per CLAUDE.md's "Skills calling skills", the fix is an input (an argument, a file section, a value passed in); the standard Hand off line is the one exception.
• Scope drift. A section that no longer fits its skill's description.

4. Fix and report

• Make the fixes on a branch, following CLAUDE.md's committing rules. Ask before committing.
• Bump the patch version if any shipped file changed, then rerun step 1.
• Report what you fixed, one line each, and what you checked and left alone.