Content
100%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.
An exemplary skill body: a gated six-step workflow with per-step 'Done when' checkpoints, fully concrete queries and code, explicit error-state handling, and strict token discipline that delegates everything not Kopai-specific to the artifact-capabilities skill. The only reference (references/recipes.md) is real, one level deep, and accurately described in the body.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and prescriptive with no padding: every sentence instructs ('Choose granularity once, here, so the window divided by it, times the number of groups, stays under the aggregate cap'), and it explicitly delegates the runtime API to the artifact-capabilities skill rather than restating it ('The runtime API... belongs to the artifact-capabilities skill... take every API detail from there'). Rationales present (e.g. 'Ranking by error count keeps a service with one failed span from topping the chart') are non-obvious domain knowledge Claude cannot infer, not over-explanation. Not 4: nothing could be trimmed without losing a real instruction or constraint. | 5 / 5 |
Actionability | Fully concrete throughout: exact tool names ('mcp__claude_ai_Kopai__query'), copy-paste-ready KopaiQuery JSON for the worked example, executable watchTool JS with the refetchInterval math spelled out ('refetchInterval equal to the query's granularity in milliseconds (5m → 300000)'), exact user-facing messages, and per-error-code handling. The undefined watchSection helper is explicitly justified ('watchSection is the page's own wrapper around mcp.watchTool: it reads result.payload.data, updates the section's "last updated" from result.cache?.storedAt, routes error events...'), satisfying the pseudocode exception. Not 4: the common case is covered end-to-end with concrete code and exact values. | 5 / 5 |
Workflow Clarity | Six sequenced steps, each with an explicit 'Done when...' validation checkpoint, plus feedback loops ('on invalid_input, repair from the result's issues and call again'; zero rows means 'fix it in step 3, never publish an empty chart') and hard gates on the risky outward-facing action ('Publish only when step 1 found Kopai', 'build if you like, then refuse to publish'). This matches the anchor-5 pattern of clear sequence, explicit validation, error-recovery loops, and checklists. Not 4: no checkpoint is implicit or missing, including the publish gate and the final checklist ('every query has a watch, every state in the list has its message, and no observed value appears in the page source'). | 5 / 5 |
Progressive Disclosure | The body is a lean overview (definition, workflow, one worked example) with a single, well-signaled, one-level-deep reference whose contents are named inline ('a recipe in references/recipes.md (latency percentiles, log volume, one metric series, a fixed set of services)'); the bundle file exists and contains exactly those recipes, and the bulk runtime API is delegated to the separate artifact-capabilities skill. Not 4: there is no content inlined in SKILL.md that belongs in a reference, and navigation is unambiguous. | 5 / 5 |
Total | 20 / 20 Passed |