Content
50%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is highly actionable — exact commands, exit-code contracts, and consent gating throughout — but it is a monolithic 567-line document suffering from verbosity and structural defects. Dream-cycle internals and version-specific migration detail belong in reference files, the Manual path's step structure is broken mid-list, and an orphaned paragraph dangles without a heading.
Suggestions
Move the dream-cycle internals (triage cascade, rescue mechanics, config-knob defaults, cooldown semantics) into a reference file such as references/dream.md, keeping only invocation patterns, consent requirements, and the quality bar in SKILL.md.
Fix the structural breaks: resume or restructure the "Manual path" numbered list so the per-dimension sections nest under step 2 instead of floating as sibling headings, and re-home the orphaned timeline-parsing paragraph (lines ~303-307) under a proper heading.
Consolidate version-specific material (v0.11.0 half-migration repair, v0.23 dream phases, v0.36.4.0 autonomous path) into a dedicated version-notes or migration section so the main workflow stays lean and date-sensitive detail is isolated.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is noticeably verbose: the dream-cycle section (~150 lines) buries invocation guidance under internal mechanics — triage cascades, rescue floors, config-knob defaults ("dream.triage.max_ms, default 5 min"), cooldown semantics, and issue IDs (#6377, #6753) — that read like release notes rather than operating instructions. Version numbers (v0.23, v0.36.4.0, v0.11.0, v0.10.3+) and dates are woven inline throughout instead of being isolated in a version-notes or deprecated section, which the rubric explicitly penalizes. It is not 1 because it never explains concepts Claude already knows — the padding is dense operational detail, not tutorial filler. | 2 / 5 |
Actionability | Mostly executable guidance: exact commands ("gbrain doctor --remediation-plan --json", "gbrain embed --stale --dry-run"), precise exit-code semantics (0/1/2 for autopilot, exit 3 for consent, exit 11 for budget), and a concrete report template. It is not 5 because several dimensions stop at direction without the command or tool call ("Remove dead links in gbrain", "Flag for review", "Standardize to the most common variant using gbrain tag operations"). | 4 / 5 |
Workflow Clarity | Sequencing and validation are genuinely present (dry-run previews, plan_hash-bound approval, cost caps, health check before/after), so the missing-validation cap does not apply — but the structure is broken. The "Manual path" numbered list is cut off after step 2, leaving the per-dimension sections floating as sibling headings, and an orphaned paragraph about timeline parsing formats dangles after the "Auto-commit deferred to v1.1" block with no heading. It is not 4 because these structural defects make the phase and step boundaries ambiguous; it is not 2 because a real, checkpointed sequence exists across the manual, autonomous, and heartbeat paths. | 3 / 5 |
Progressive Disclosure | The many ## and ### headers make the 567-line body navigable (not 2's "no section headers"), but content that clearly belongs in separate files is inlined: the dream-cycle internals, the half-migrated-install repair, and the heartbeat schedules each warrant their own reference file. References to external docs ("docs/guides/repair.md", "skills/minion-orchestrator/SKILL.md", "../conventions/agent-operator-protocol.md") are clearly signaled but none exist in the bundle — there is no references/ directory — so the skill is a monolith rather than an overview with one-level-deep references. It is not 4 because the bulk of specialized detail sits inline in SKILL.md instead of being split out. | 3 / 5 |
Total | 12 / 20 Passed |