NightMarquee presents three well-named tools with clear, descriptive documentation and proper schema definitions. Tool names follow verb_noun conventions and are action-oriented (search, get, sign_in). Descriptions are detailed and contextual, ranging 150-250 characters, well above the 10-char minimum and within the optimal 50-200 range for LLM interpretation. Input schemas are fully defined with JSON Schema types, descriptions, enums, and constraints. However, the server lacks explicit error recovery guidance (friendly() helper exists but not exposed to LLM), does not implement confirmation/dry-run for the destructive sign_in tool, and output schemas are declared but not fully documented in the definitions. The composition is clean, three single-responsibility tools with clear chaining (search → get) and proper parameter naming alignment. Indexed descriptions are informative for the LLM decision-making process.

• Add explicit input schema and description to nightmarquee_sign_in. Document that it takes no parameters, initiates a device-code OAuth flow, and saves credentials locally. Include note: 'No parameters required. Follow the device code shown in the output to authorize on nightmarquee.com.'
• Implement and expose error classification in tool responses. When apiGet() or device OAuth fails, return structured error with category ('retryable', 'user_fixable', 'fatal'), actionable next step, and any context (e.g., 'Account does not have Unlimited plan, see https://nightmarquee.com/pricing'). Reference pattern:error-classification.
• Formally declare output schemas in tool definitions using outputSchema property. For nightmarquee_search_prompts, document the response structure: { total: number, results: Array<{slug, title, tagline, category, tier, demoUrl}> }. For nightmarquee_get_prompt, document: { slug, entitled, tier, title, tool, text?, demoUrl, pricingUrl? }.
• Clarify default behavior for the 'brand' parameter in nightmarquee_get_prompt. Update description: 'Swap the fictional brand for this name (max 24 chars). If omitted, uses the original brand from the prompt.'
• Add a dry-run or confirmation step for nightmarquee_sign_in, or explicitly document that calling it multiple times is safe (idempotent) and does not consume OAuth tokens. This guides agents on whether they can safely retry on connection errors.
• Validate and document the 'limit' parameter bound (1-50) in nightmarquee_search_prompts. Current schema has minimum and maximum constraints visible in code; ensure they are enforced at tool invocation and that oversized requests return a helpful error like 'limit must be 1-50; requested 100'.