Content
71%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.
The body is highly actionable with executable patterns throughout and a clear intent-to-shape decision workflow, but it is over-long for a single file: rules are duplicated, changelog detail is inlined, and reference-grade sections (language ports, MCP event routing, playbook evolve) belong in separate reference files. Splitting those out and deduplicating the discovery/child-agent rules would materially raise both conciseness and progressive disclosure.
Suggestions
Move reference-grade detail out of SKILL.md into bundled reference files — the generated-language observer/evolve table, MCP event-source/webhook mapping rules, and playbook construction-time learning internals — and link them with clearly signaled one-level-deep references.
Deduplicate repeated rules: the discovery guidance ("If discovery is enabled, call discover(...) before using callables") appears in both Use These Defaults and Critical Rules, and the child-agents-in-functions rule is stated in Use These Defaults, Critical Rules, and Child Agents As Tools; state each once and cross-reference the section.
Trim changelog-style asides that do not affect code generation, e.g. "Enabled by default after its A/B gate passed on both small and frontier models and implemented in the generated language ports through AxIR Core" and the `seed` snapshot deprecation history, keeping only the behavior and the opt-out.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes competence, but it repeats rules ("If discovery is enabled, call discover(...)" appears in both Use These Defaults and Critical Rules; child agents in "functions: [...]" is stated three times) and carries changelog-style asides such as "Enabled by default after its A/B gate passed on both small and frontier models" and "their older `seed` snapshot key still works with a deprecation warning". This is more than the 'minor instances' of a score-4 anchor but not the pervasive padding of a 2. | 3 / 5 |
Actionability | Nearly every section ships copy-paste-ready TypeScript: the canonical agent() pattern, audio I/O, child agents as namespaced tools, fn(...) builders, grouped modules, extra.protocol.final(...) handlers, the full clarification/resume try-catch flow, and bubbleErrors. The Decision Guide maps concrete user intents to specific agent shapes, covering the common cases. | 5 / 5 |
Workflow Clarity | The Decision Guide plus Do Not Generate section sequence the work clearly (intent -> shape -> code), and the clarification/resume pattern is a full catch -> error.getState() -> agent.setState(savedState) -> forward() recovery loop. It is not a 5 because there is no explicit validation checkpoint for generated output, though this codegen skill involves no destructive or batch operations that would mandate one. | 4 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/), so all ~700 lines are inline. Headers are clear and the skill defers many topics to sibling skills, but reference-grade content sits inline in SKILL.md: the generated-language observer/evolve table, MCP event-source and webhook mapping rules (a section that itself says "Keep this section focused on Agent attachment"), and playbook internals. A 4 would require most of that detail to live in separate files with clearly signaled references. | 3 / 5 |
Total | 15 / 20 Passed |