Content
86%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 lean, executable integration guide — complete code, real error-handling checkpoints, and high-signal pitfall notes with no basic-concept padding. The weaknesses are a dangling 'use-topbar' reference in a bundle with no reference files and the absence of an explicit enforcement/verification step for the batch-completion cap.
Suggestions
Resolve the 'use-topbar' reference: either add the referenced file under references/ and link it as 'See [use-topbar.md](references/use-topbar.md)', or inline the one or two required Topbar configuration lines so Step 3 is self-contained.
Add an explicit validation checkpoint to the batch-LLM hard gate, e.g. 'Before finishing, count completion call sites in generated code and confirm <= 5 per user action and 50 hard ceiling' — this turns the cap table into a verify step rather than guidance only.
Consider moving the three full agent-server source files into a references/ code template (with the body keeping signatures and key lines), which would cut body tokens and resolve the skill's only structural gap.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Dense imperative one-liners with zero padding ('Register on mount, unregister on unmount'); it explains only non-obvious SDK/React pitfalls like 'Comlink proxies are callable — setApi(proxy) ... stores a Promise' that Claude could not infer. Every code block earns its tokens. | 5 / 5 |
Actionability | Fully executable end to end: 'pnpm add @cognite/app-sdk' with minimum version, a file layout, four complete copy-paste TypeScript files (hook, resources, actions, server wiring), and a runnable unit-test snippet covering the factories. Not 4 because there are no pseudocode gaps or missing key details. | 5 / 5 |
Workflow Clarity | Steps 0–6 are clearly sequenced with real checkpoints: catch the outside-Fusion rejection, '.catch()' on register/unregister, and a final checklist. Not 5 because the batch-LLM hard gate states caps (5/50) and UX disclosure but no explicit validation step to verify the cap is enforced before shipping, so a feedback loop is missing for the batch case. | 4 / 5 |
Progressive Disclosure | Step headers are well organized, but 'use-topbar' is referenced twice (Step 3 'see use-topbar' and the Checklist) as a bare inline code mention with no link, and no references/ or other bundle directory exists in this skill — the reference is dangling and unclearly signaled. This matches the 3 anchor (some structure, references present but not clearly signaled) rather than 4, where references would be mostly clear and resolvable. | 3 / 5 |
Total | 17 / 20 Passed |