Content
85%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 well-crafted instruction-only skill: exact output formats, two operational decision rubrics, and a procedure with a mandatory human validation gate before the irreversible publish step. The two 4-scores reflect polish gaps — some persuasive padding in the rules/anti-patterns, and no worked example ticket — rather than structural problems.
Suggestions
Add one worked example ticket (title + full body with Blocked by/Tier/Domain lines and the ready-for-agent label where applicable) so the output artifact is copy-paste exemplified rather than only described — this closes the actionability gap.
Trim the justificatory asides in the rules and anti-patterns (e.g. rule 9's incentive explanation, "wearing a different label", "the second is just 'no decision' with extra words") down to their operative sentences; the rationale is useful but currently costs tokens on every load.
In the tier rubric, the pointer to "this repo's local workflow article" is the least concrete reference in the body — either name the file path (as done for scripts/agents.config.sh) or drop the pointer.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes competence — it cites "shared invariant §2/§4/§6/§10" and root AGENTS.md rules without re-explaining them — and it never explains concepts Claude already knows. It sits at 4 rather than 5 because several passages are persuasive rationale rather than instruction, e.g. rule 9's "an agent asked to size itself has every incentive to answer 'the strongest one'" and anti-patterns' "the second is just 'no decision' with extra words", which could be trimmed without losing guidance. | 4 / 5 |
Actionability | Guidance is highly concrete for an instruction-only skill: exact stamp formats ("Blocked by: #N", "Tier: <planner|implementer|mechanical|reviewer>", "Domain: <token>" with regex "[a-z][a-z0-9-]*"), a real command ("gh issue create"), label names, and two decision rubrics with worked examples ("A rename across call sites, a codemod, a dependency bump"). It misses the 5 anchor only because no assembled example ticket (title + body) is shown, leaving the output artifact described rather than exemplified. | 4 / 5 |
Workflow Clarity | The procedure is a clear 5-step sequence (read PRD → draft → quiz → publish → hand off) with an explicit mandatory validation gate before the risky batch action: "Quiz step (mandatory human gate)... Do not publish until they confirm." The quiz functions as a full feedback loop — the user challenges granularity, ordering, labels, and tiers, and tier uniformity is surfaced as a finding — and the anti-patterns section enumerates failure modes. | 5 / 5 |
Progressive Disclosure | No bundle files exist, and the skill appropriately delegates detail to clearly-signaled repo files one level deep — tier definitions to "the root AGENTS.md", tier→model mappings to "scripts/agents.config.sh", domain vocabulary to the same config, terms to "docs/domain-glossary.md" — rather than duplicating or nesting them. The remaining content is a single well-sectioned file (rules, two rubrics, trust boundary, procedure, anti-patterns) where every section is needed at decision time. | 5 / 5 |
Total | 18 / 20 Passed |