Strong tool naming (all verb-noun), comprehensive descriptions (150-250 chars), and well-structured schemas with enums and constraints. All 7 tools have explicit registrations with inputSchema, descriptions, and annotations. Parameters are typed and described. Main gaps: output schemas are not formally documented (only inferred from respond() wrapper), error messages are generic, and no pagination guidance for large result sets. Composition is clean, each tool has one responsibility. Rate limiting and telemetry are present but not exposed to the LLM.

• Add formal outputSchema to each tool registration. Example: search_fragrances returns {result_count: number, fragrances: [{name, brand, family, price_usd, notes, accords, ...}], attribution: string}. This enables LLMs to plan downstream operations.
• Enhance error handling with recovery hints. When mustResolve() fails, return: 'Fragrance "X" not found. Try search_fragrances(query="X") to find similar names, or browse trending_fragrances().'
• Document valid values for get_recommendations.preferences. Either provide an enum of ~50 common notes/accords, or add a discover_notes tool that lists available options.
• Add total_count and next_cursor to search_fragrances and trending_fragrances responses. This enables agents to paginate large result sets without context explosion.
• Clarify slug vs name in tool descriptions. Add: 'Accepts either a Perfume Picks slug (e.g., bleu-de-chanel) or a fragrance name (e.g., Bleu de Chanel). If you have only a name, call search_fragrances first to get the slug.'