Conciseness

The body is dense and operational (tables for stages, statuses, and severity; terse rules like 'Never prune what a model cannot absorb'), but the opening conceptual framing — 'An agent without a flywheel is limited to a single session: what it learns ends when the session does' and 'Every harness records sessions and stores context differently' — is motivational context that could be trimmed. This matches the 4 anchor (efficient with minor instances that could be trimmed) rather than 5 (every token earns its place); the Katz citation does earn its place since the concept is niche.

Actionability

Guidance is fully executable: exact discovery command (`git ls-files ':(glob)**/*flywheel*.md' ...`), exact evidence commands (`git log -1 --format=%cs -- <path>`, `python3 ~/scripts/harness-sessions.py <slugs> --since <N> --json`, `command -v`, `just --list`), and copy-paste templates for the review report, Cycle log row, pointer section, and charter row. This matches the 5 anchor; nothing is pseudocode or left at the hint level.

Workflow Clarity

Both modes are numbered procedures with explicit validation and feedback loops: review step 2 specifies path/command/hook/recency checks, step 6 gates all changes on a user answer ('Change nothing before the answer'), and step 8 reruns the checks after repair. Destructive operations are explicitly guarded ('Never remove a scaffold without an ablation result', 'Never overwrite a flywheel doc', never edit permission-defining files — write a patch instead). This matches the 5 anchor's explicit validation steps and error-recovery loops.

Progressive Disclosure

All referenced bundle files exist (references/stages.md, references/harness-signals.md, references/setup-prompts.md, assets/flywheel-template.md), are one level deep, and each reference is signaled with its purpose and read timing ('Read it before judging or proposing any stage'). The split is appropriate: procedures and report formats live in SKILL.md while per-stage evidence checks, harness specifics, and setup prompts live in references — matching the 5 anchor's clear overview with well-signaled one-level-deep references.

Specificity

The description lists multiple concrete actions covering the full loop — 'Define, review, repair, or set up', 'records agent sessions', 'learns from them by analysis or by replaying them on better models', 'applies the lessons to instructions, skills, hooks, and memory', 'prunes scaffolding', 'tunes the loop itself' — with no vague filler. Coverage is comprehensive across all six capabilities, matching the 5 anchor rather than the 4 anchor (which expects minor gaps).

Completeness

It explicitly answers 'what' (define, review, repair, or set up the six-stage learning loop, each stage named as an action) and 'when' via a concrete 'Use when' clause with three trigger scenarios (existing doc needing review/repair, no learning loop yet, user questions about agent learning, memory, retros, evals, or scaffold removal). This matches the 5 anchor's structure exactly; the 4 anchor's 'when could be more explicit' does not apply.

Trigger Term Quality

Natural trigger terms include 'flywheel', 'learning loop', 'docs/flywheel.md', 'review or repair', 'run retros or evals on recorded sessions', 'keep memory', and 'remove harness scaffolding', covering synonyms and the concrete file path. This is at least as comprehensive as the 5-anchor example ('PDF files, PDFs, forms, document extraction, .pdf'), not merely 'good coverage with a few missing' as at 4.

Distinctiveness Conflict Risk

It occupies a clear niche — a repo's agent harness data flywheel — with distinct triggers unlikely to fire for unrelated skills; even the broadest phrases ('keep memory', 'run retros') are qualified by the agent-learning-loop context. The 4 anchor's 'minor overlap risk with closely related skills' is not a better fit since the triggers and domain are specific.