Conciseness

The body avoids tutorial prose and assumes competence ("Prefer short, modern, copyable patterns. Do not write tutorial prose"), and most lines are library-specific rules Claude cannot know (e.g. "noul >= trueThreshold", the sampling-parameter matrix). But at ~1,090 lines it is padded with repetition ("Do not confuse the `amazon-bedrock` OpenAI-compatible profile" reappears under Critical Rules and Do Not Generate; WebLLM/Chrome-AI portability rules restated) and includes a sentence truncated mid-clause ("a stream that has started runs on"), so it could be noticeably tightened.

Actionability

Nearly every section opens with copy-paste-ready TypeScript using real identifiers ("const llm = AxBalancer.create([openai, anthropic] as const, { strategy: { type: 'adaptive', ... } })", "await gemini.speak({ text: 'Hello from Ax.', voice: 'Kore' })"). Minor gaps keep it below fully-executable: some examples reference undefined context variables (`gen`, `values`, `base64Wav`, `apiKey`) and the Common Options bullet is cut off mid-sentence.

Workflow Clarity

Some sequencing exists ("Use axGetSupportedAIModels() to build provider/model selectors before creating an ai(...) instance", "Choose the primitive by responsibility") and decision tables (budget levels) order the choices, but there is no explicit end-to-end workflow (select profile → configure → call → verify) and no validation checkpoints, so the sequence is present but implicit. No destructive/batch operations, so no cap applies.

Progressive Disclosure

Sections are clearly headed and external references are well-signaled and one level deep ("Fetch these for full working code" linking 18 example files, plus the ax-typesafe/signature skill links), but no references/, scripts/, or assets/ bundle exists — provider-specific minutiae (Meta Muse, Bedrock, Chrome AI, WebLLM, the DeepSeek notes) that clearly belong in separate reference files are all inlined in a single ~1,090-line SKILL.md.

Specificity

"generate correct AI provider setup and configuration code using @ax-llm/ax" states concrete actions, and the trigger list enumerates specific capability areas ("routing, adaptive balancing, presets, embeddings, batch audio with ai.transcribe() or ai.speak(), extended thinking, context caching") — several specific items, though mostly topic nouns rather than the multiple concrete verb-actions of a 5.

Completeness

Both parts are explicit: what — "generate correct AI provider setup and configuration code using @ax-llm/ax"; when — "Use when the user asks about ai(), providers, models, routing, ... or mentions OpenAI/Anthropic/Google/Azure/DeepSeek/... with @ax-llm/ax", with concrete trigger phrases throughout.

Trigger Term Quality

Good natural-term coverage: API names users would say ("ai()", "ai.transcribe()", "ai.speak()"), capability nouns ("providers, models, routing, embeddings, context caching, extended thinking"), ten provider names, and the package name itself. A few common phrases are missing (e.g. "chat", "streaming", "API key"), keeping it below comprehensive-with-synonyms.

Distinctiveness Conflict Risk

The @ax-llm/ax qualifier and library-specific triggers ("ai()", "Typesafe/Jev") carve a clear niche, but generic terms like "providers, models, embeddings, routing" overlap with other AI-SDK skills unless the library is mentioned — minor overlap risk with closely related skills.