Content
82%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 lean, well-structured command body with a concrete executable decision-engine invocation, explicit kill-criteria checkpoints, specific output expectations, and useful anti-patterns. The main gaps are missing error-recovery/merge guidance in the workflow, specialist forks lacking an invocation mechanism, and core references that live outside the skill bundle and cannot be verified to exist.
Suggestions
Ship the referenced files inside the skill bundle (e.g. references/forcing_questions.md, scripts/frontend_decision_engine.py, profiles/*.json) or use bundle-relative paths so the core protocol is self-contained and verifiable.
Add error-recovery guidance to the routing protocol: what to do when the decision engine script fails or specialist findings conflict with the matched profile, and when/how the depth-first specialist fork terminates and merges results.
Make the specialist fork step concrete by specifying the invocation mechanism (e.g. the Agent/subagent call or command) and define the runner-up tradeoff threshold precisely (e.g. 'within 15% match-score of the top profile').
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is ~60 lines of dense lists, commands, and bullets with zero padding: it never explains what LCP, WCAG, or design systems are, and every section (forcing questions, routing protocol, output expectations, anti-patterns, customization) carries purely operational content. This matches the 5 anchor (lean, assumes Claude's competence, every token earns its place); there is no over-explanation to trim that would justify 4. | 5 / 5 |
Actionability | Concrete, executable guidance dominates: a full bash invocation with enumerated flag choices ("--primary-device <mobile-4g|desktop-fiber|low-end-android|corporate-network>", "--lcp-target-ms <N>"), named fork conditions ("epic-design only for astro-or-static marketing surfaces"), and measurable outputs ("p75 on the primary device", "KB-gzip"). Not 5 because of minor gaps: the referenced script and canonical reference file do not exist in the skill's bundle, the specialist-fork step gives no concrete invocation mechanism, and "runner-up tradeoff (if within 15%)" does not define 15% of what; not 3 because the command block is directly executable with fully specified parameters. | 4 / 5 |
Workflow Clarity | The routing protocol is a clearly numbered 5-step sequence with an explicit validation checkpoint ("Surface kill criteria — e.g., 'SEO-dependent + SPA-only' trips. STOP and resolve.") and a concrete final checklist ("Output expectations (≤ 200-word digest)"), matching the 4 anchor. Not 5 because there is no error-recovery feedback loop (e.g., what to do when the decision engine fails or specialist findings conflict) and the depth-first fork loop has no completion or merge criteria; not 3 because checkpoints are explicit and the sequence is unambiguous. | 4 / 5 |
Progressive Disclosure | The body is well-organized into clear sections, keeps an appropriately brief inline 7-question summary while deferring detail to a clearly labeled canonical reference ("Canonical source: engineering-team/skills/senior-frontend/references/forcing_questions.md"), and signals every external resource with a full path and purpose. Not 5 because the referenced paths (forcing_questions.md, frontend_decision_engine.py, profiles/) point outside the skill bundle and do not resolve — no references/, scripts/, or assets/ directories ship with the skill — so the core protocol depends on unverifiable external files; not 3 because signaling is clear and nothing that should be separate is inlined. | 4 / 5 |
Total | 17 / 20 Passed |