Strong foundation with complete schemas, clear descriptions, and proper input validation via Zod. All 10 tools have explicit registrations with descriptions (avg 95 chars) and typed parameters. Naming follows verb_noun convention (add_lead, qualify_lead, move_stage). However, output schemas are undocumented, responses are text-wrapped without formal structure declarations. Error messages lack recovery guidance (e.g., 'CRM operation failed' is generic). No tool annotations (readOnlyHint/destructiveHint) despite clear WRITE vs READ_ONLY risk classifications. Parameter descriptions are solid but could be more prescriptive about constraints (e.g., 'Lead ID matching pattern L-\d{4,}' is stated but not in description text for all tools).

• Document output schemas for all tools. E.g., add_lead should declare: {type: 'object', properties: {id: {type: 'string'}, name: {type: 'string'}, ...}, required: ['id', 'name']}. This enables LLMs to chain calls and extract typed fields.
• Add tool annotations: mark add_lead, qualify_lead, move_stage, log_activity, import_csv, trigger_n8n with destructiveHint: true; mark list_pipeline, due_followups, search_leads, export_csv with readOnlyHint: true. This signals safety constraints to clients.
• Replace generic error messages with actionable recovery guidance. E.g., 'Lead L-0001 not found. Try search_leads(query="name") to find the correct ID.' or 'Invalid score: must be 0-100, got 150.'
• Implement pagination for list_pipeline and search_leads: add next_cursor to responses and accept cursor parameter. Cap default limit at 20 and document in descriptions.
• Add dry-run support to import_csv and trigger_n8n. E.g., import_csv(path, dry_run=true) returns preview of changes without committing. trigger_n8n already has dry-run logic but should expose it as a parameter.