Content
63%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.
A rich, concrete workflow document with strong sequencing, explicit user-approval checkpoints, and well-specified gate feedback loops. Its main costs are runtime-irrelevant editorial commentary, triple-repeated gate-handling boilerplate that belongs in one referenced section, and a Parse Arguments section that under-specifies the argument modes the rest of the skill uses.
Suggestions
Factor the three near-identical review-mode-check blocks and the three NOT ASSESSED handling paragraphs into one shared instruction (e.g. a short "Director gates & review modes" section) referenced from each phase.
Remove or relocate the editorial blockquotes explaining why lines exist ("The loop needed an exit…", "Keep this line scoped to its mode") — they are author-facing review notes, not runtime instructions.
Fix Parse Arguments to actually cover all three entry modes (`next`, `[system-name]`, bare) and the `--review full|lean|solo` override, matching the argument-hint and Phase 6.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The core guidance (inference patterns, dependency heuristics, tier assignment, gate handling) is project-specific and earns its place, but the body carries ~40 lines of editorial blockquotes justifying past edits ("The loop needed an exit that does not depend on being asked", "Keep this line scoped to its mode") and repeats the review-mode-check plus NOT ASSESSED handling protocol three times nearly verbatim. Mostly efficient with clear tightening opportunities — anchor 3, not 2, because the padding is meta-commentary rather than explanations of concepts Claude already knows. | 3 / 5 |
Actionability | Highly concrete for an instruction-only skill: exact AskUserQuestion strings, exact file paths, named gates with the fields to pass ("the dependency graph (each system → what it depends on), layer assignments, bottleneck systems…"), worked inference patterns ("Inventory" implies: item database, equipment slots…), and exact verdict/note formats. Not 5 because Step 5a delegates the actual document creation entirely to an unseen template, and the Parse Arguments section promises modes it does not specify. | 4 / 5 |
Workflow Clarity | Phases 1–7 are explicitly sequenced with approval checkpoints at each collaborative step and genuine feedback loops for error recovery (REJECT → revise with user → re-run gate; NOT ASSESSED → name missing input → supply and re-run). Not 5 because the opening before "Parse Arguments" is fragmented (hook-output handling and an orphan reference to director-gates.md appear before any context), and "Two modes" contradicts the three entry modes Phase 6 actually handles ([system-name], next, bare invocation), leaving the `--review` flag's parsing undefined. | 4 / 5 |
Progressive Disclosure | Section structure is clear and external references are explicitly pathed (gate files correctly delegated: "the spawned agent reads its own gate file; do not read it in the parent session"), but the ~450-line monolithic body inlines substantial content that belongs in a referenced doc — the review-mode-check block and NOT ASSESSED gate-handling protocol each repeated three times, plus the editorial column-contract analysis. Anchor 3 (content that should be separate is inline) fits better than 4 given the volume of factored-able boilerplate. | 3 / 5 |
Total | 14 / 20 Passed |