The body is dense and assumes Claude's competence — it never explains git, worktrees, or basic refactoring — and each section (ritual, gate, stop conditions) earns its place without padding.

Provides fully executable commands (`rg` caller counts, `bun test`, `bun run typecheck`, `git worktree add`, `git diff --name-only`) plus an exact copy-paste finding-format template and commit citation requirement.

The per-iteration ritual is a clear numbered sequence with explicit validation checkpoints (anti-cosmetic gate, test/typecheck), a re-grep feedback loop for stragglers, and enumerated stop conditions.

The SKILL.md body is an overview that signals six one-level-deep references via 'Load on demand' / 'For X, read Y' phrasing, and all referenced files (never-touch, smell-catalog, library-refusal, report-format, goal-template, periodic-monorepo-pass) exist in references/.

Lists multiple concrete actions — 'surgically removes indirection failing to earn its boundary', 'audit a package for dead abstractions', and 'a sequence of small refactor commits that delete more than they add' — matching the score-3 anchor for multiple specific concrete actions.

Clearly answers both what ('Run a continuous collapse-and-simplify pass that surgically removes indirection') and when via an explicit 'Use when...' clause with multiple triggers.

Surfaces natural phrases a user would actually say — 'collapse pass', 'simplify pass', 'reduce indirection', 'shrink the surface', 'find what to delete' — giving good coverage of common variations.

Occupies a clear niche with distinctive trigger phrases ('collapse pass', 'simplify pass') unlikely to fire for unrelated skills, so it does not hit the level-2 overlap concern.