Lean throughout — no explanations of concepts Claude already knows; the API-surface list and Gotchas are dense and each token earns its place (e.g. the inline 'Swap this for a mic capture' comment is guidance, not filler).

Three copy-paste-ready quick-starts (REST URL, REST file, WebSocket) with real model/language params and exact method calls, plus a concrete list of WSS client messages and server events.

The WSS two-stage flow is explicitly sequenced ('createConnection() → register handlers → connect() → waitForOpen()') and the Gotchas act as checkpoints; transcription is non-destructive so no validate/feedback loop is required.

No bundle files exist, so judged on inline structure and referenced paths; the 'API reference (layered)' section points one-level-deep to canonical OpenAPI/AsyncAPI/docs and in-repo src paths instead of reproducing them, with example files clearly listed and well-organized sections.

Lists multiple concrete actions and exact SDK method names — `transcribeUrl` / `transcribeFile` (REST) and `createConnection()` / `connect()` (WebSocket) — rather than vague language.

Opens with an explicit 'Use when...' clause (when) and states the covered REST + WebSocket surface (what), then enumerates triggers — both halves are explicit.

Provides a broad trigger list a developer would naturally say ('transcribe', 'speech to text', 'STT', 'live transcription', 'websocket transcription') alongside product-specific terms ('nova-3', 'listen.v1').

Scoped tightly to Deepgram v1 STT in JS/TS and actively disambiguates by redirecting summarize/diarize, Flux turn-taking, and full-duplex assistants to three named sibling skills.