Conciseness

Mostly efficient, genuinely Neon-specific material (env var semantics, dialect routes, plan gating) but with clear tightening opportunities: the region list appears verbatim twice ("Currently available in aws-us-east-2..." at the top and again under Availability), the dialect path routing is explained in full twice (the routes list plus 'So ${NEON_AI_GATEWAY_BASE_URL}/v1 is the chat-completions endpoint...' and again in 'Use with Plain SDKs'), credential sourcing is explained three times (Setup, Environment Variables, and the three-bullet 'Getting the credentials' block), and there is marketing prose ("runs on the same Databricks infrastructure that serves trillions of tokens a month", "batteries-included"). Not 2: it never explains concepts Claude already knows, and the padding is a fraction of useful content; not 4: the duplication is repeated and prominent rather than minor.

Actionability

Fully concrete, executable guidance throughout: the neon.ts config block, `neon deploy` / `neon config status|plan|apply` commands, the exact env var table (`NEON_AI_GATEWAY_TOKEN`, `NEON_AI_GATEWAY_BASE_URL` with the bare-host format spelled out), copy-paste TypeScript for the AI SDK provider (streamText, generateText, agent with tools/stopWhen), Mastra, and plain OpenAI SDK in both dialects, plus a runnable curl for /v1/models with the full response shape. The common cases are all covered specifically; not 4 because the only gaps are illustrative undefined bindings (messages, db, memory) that do not reduce the specificity of the instructions.

Workflow Clarity

The sequence is clear and well-ordered — preconditions (region, paid plan), enable via neon.ts, `neon deploy`, credentials pulled into .env.local, then SDK usage — with explicit checkpoints (`neon config status` / `neon config plan`, reading /v1/models) and an error-recovery distinction for the free-plan-blocked vs reduced-catalog cases. Not 5: there is no post-deploy validation that a first inference request actually succeeds, and several examples rely on implicitly available bindings (messages, db, memory) rather than a checked step; not 3: validation and debugging guidance are explicit, not implicit.

Progressive Disclosure

No bundle files (references/, scripts/, assets/) exist; everything is inline in a ~295-line body. Structure is good — clear sections in dependency order, well-signaled one-level-deep external pointers (the parent neon skill, the neon-functions skill's ai-sdk.md, models.dev, and a Further Reading list of doc URLs). Not 5: reference-worthy detail is inlined rather than split — the full ~30-line /v1/models response JSON, the four-dialect routing table, and the plan-gating detail would each fit a reference file; not 3: navigation is easy, references are clearly labeled, and no content is buried or nested.

Specificity

Concrete actions are named — "call an LLM", "route between model providers (OpenAI, Anthropic, Google/Gemini, Meta, Alibaba, and more)", "avoid juggling separate provider API keys", "Works with the OpenAI SDK, Anthropic SDK, google-genai, the Vercel AI SDK, and Mastra by changing only the base URL" — which reaches anchor 5's breadth, but the second-person phrasing "built into your Neon branch" triggers the rubric's voice penalty, reducing it to 4. Not anchor 3: far more than 1-2 concrete actions; not clean 5: penalized voice plus minor coverage gaps (no mention of model listing or streaming as actions).

Completeness

Explicitly answers both: what — "One API and one credential for frontier and open-source LLMs, built into your Neon branch", "Works with the OpenAI SDK, Anthropic SDK, google-genai, the Vercel AI SDK, and Mastra by changing only the base URL" — and when — "Use when a user wants to call an LLM, add AI/chat/an agent to their app, route between model providers... or avoid juggling separate provider API keys and accounts", with concrete trigger phrases enumerated. Matches anchor 5 exactly; anchor 4's 'when could be more explicit' does not apply.

Trigger Term Quality

Trigger coverage is broad and natural — "call an LLM", "add AI to my app", "chat completion", "model routing", "LLM proxy/gateway", "one API for all models", "use Claude/GPT/Gemini", "Neon AI Gateway" — but the listed trigger "log/rate-limit AI calls" is an over-claim: neither the description's capabilities nor the skill body mention logging or rate limiting, so the trigger is unsupported. Not 5 due to that unsupported term; not 3 because the rest of the coverage, including synonyms and product name, is comprehensive.

Distinctiveness Conflict Risk

The Neon-specific framing ("Neon AI Gateway", "built into your Neon branch", one Neon credential) carves a clear niche, but broad triggers like "call an LLM", "add AI to my app", "use Claude/GPT/Gemini", and "AI SDK" overlap with generic LLM-provider skills and would fire for users with no Neon context. Mostly distinct with minor overlap risk — anchor 4; not 5 because of that generic-trigger overlap, not 3 because the core subject is unambiguous.