Maverick MCP has significant definition quality gaps. Tool names lack clear action verbs (maverick_assess, maverick_profile, maverick_adapt are domain-specific but not verb-noun patterns). Descriptions are present but generic, most under 200 chars and lack WHEN/WHY context. Input schemas exist but are minimal: maverick_assess has an empty 'properties' object for 'answers', maverick_profile has empty input schema, and parameter descriptions are sparse. No output schemas documented. Error handling is minimal, responses use generic 'status' fields without recovery guidance. The server reads/writes local YAML files without validation or permission checks. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear read/write distinctions.

• Rename tools to verb_noun pattern: run_assessment, get_profile, get_adaptation, generate_card, record_feedback. This matches LLM expectations and arcade.dev/patterns/tool baseline.
• Add complete input schemas with typed properties and enums. For maverick_assess.answers, define: {type: 'object', properties: {1: {type: 'string', enum: ['a', 'b']}, 2: {...}}, required: []}. For maverick_profile, define empty input as {type: 'object', properties: {}, required: []}.
• Document output schemas for all tools. Example for maverick_profile: {type: 'object', properties: {status: {type: 'string', enum: ['ok', 'no_profile', 'error']}, profile_id: {type: 'string'}, name: {type: 'string'}, drives: {type: 'object', properties: {dominance: {type: 'integer', minimum: 1, maximum: 10}, ...}}, ...}}.
• Add tool annotations in MCP registration: readOnlyHint=true for maverick_profile and maverick_adapt; destructiveHint=true for maverick_assess, maverick_card, maverick_feedback (they write to ~/.maverick/).
• Enhance error responses with recovery guidance. Instead of {status: 'no_profile', message: '...'}, return {status: 'no_profile', message: '...', next_step: 'Call run_assessment to create a profile', recovery_tool: 'run_assessment'}.