No filler or explanation of concepts Claude already knows; every section adds non-obvious domain knowledge (the +AND+ pitfall, token-vs-phrase matching, MedDRA/VeDDRA spelling quirks, verified hit totals), and repeated surveillance caveats are deliberate safety reinforcement rather than padding.

Provides copy-paste-ready tool invocations with real verified outputs (e.g. OpenFDA_search_device_adverse_events with meta.results.total = 16619) plus exact tool names, endpoints, and Lucene syntax.

A clear 7-step sequence with explicit gates: a query-grammar preflight checklist (avoid special characters, do not use +AND+) and an out-of-scope stop-and-route checkpoint; this is read-only retrieval so the destructive/batch validation cap does not apply.

The body is a well-organized overview that signals exactly one one-level-deep reference ('See references/openfda_fields.md for the full per-endpoint field reference'), and that bundle file exists and holds the detailed per-endpoint field list.

Lists multiple concrete actions and named endpoints (MAUDE device adverse events, device recalls + 510(k), CAERS, veterinary, drug shortages, enforcement/recall) rather than vague language.

Clearly answers both what (multi-product post-market safety retrieval across specific endpoints) and when (explicit USE WHEN triggers), so it is not the level-2 'when only implied' anchor.

The 'USE WHEN the user asks:' clause covers many natural phrasings users would actually say (e.g. 'is [drug] in shortage', 'device recalls for [firm]', 'CAERS reactions for [brand]').

Explicit 'NOT covered by the drug-AE signal skills' carve-out plus a 'DO NOT USE for drug adverse-event SIGNAL detection' routing clause give it a clear niche unlikely to trigger the wrong skill.