Conciseness

The body is two lines — a provenance comment ("Generated by scripts/agent-harnesses.py; do not edit directly") and a single redirect instruction — with zero padding and no explanation of concepts Claude already knows. Every token serves a purpose, matching the lean score-5 anchor; there is nothing to trim, so score 4 does not apply.

Actionability

"Load and follow the shared skill at `~/.agents/skills/flywheel/SKILL.md`" is a concrete, immediately executable instruction with a specific path, so it clears the mostly-executable score-4 anchor. It falls short of score 5 because all substantive guidance is delegated to an external file with no in-bundle content, no coverage of common cases, and no fallback if that path is absent.

Workflow Clarity

This is a single-task skill and its one action — load and follow the named shared skill file — is unambiguous, which per the simple-skill guideline earns a 5. No destructive or batch operations are involved, so the validation-cap rule is not triggered.

Progressive Disclosure

The single reference to `~/.agents/skills/flywheel/SKILL.md` is clearly signaled and one level deep with nothing inlined that belongs elsewhere, giving good structure. Minor gaps keep it from a 5: the referenced path is external to the bundle and unverifiable from it, and the comment cites `scripts/agent-harnesses.py` although no `scripts/` directory exists in the bundle.

Specificity

The description enumerates multiple concrete actions — "Define, review, repair, or set up", "records agent sessions", "replaying them on better models", "applies the lessons to instructions, skills, hooks, and memory", "prunes scaffolding", "tunes the loop itself" — giving comprehensive coverage of the skill's capabilities. It matches the score-5 anchor (multiple specific concrete actions) and exceeds score 4, which anticipates gaps in coverage.

Completeness

It explicitly answers "what" (define/review/repair/set up the flywheel learning loop with its record→learn→apply→prune→tune stages) and "when" via an explicit "Use when..." clause with three concrete scenarios. This is the clear score-5 pattern; the when-clause is present and specific, so the score-4 anchor (when could be more explicit) does not fit.

Trigger Term Quality

Natural trigger phrases a user would actually say are well covered: "flywheel", "learning loop", "review or repair", "run retros or evals on recorded sessions", "keep memory", "remove harness scaffolding", plus the concrete file path "docs/flywheel.md". This matches the comprehensive-synonyms anchor; it is above score 4 because both informal terms ("retros") and file-level markers are present.

Distinctiveness Conflict Risk

The agent-harness data flywheel is a clear niche with distinct triggers (flywheel docs, learning loop, retros/evals on recorded sessions, scaffolding pruning), posing minimal conflict risk with other skills. Triggers like "keep memory" are scoped by the surrounding agent-learning context, so the minor-overlap language of the score-4 anchor is not a better fit.