All 9 tools have basic descriptions (10-50 chars) and simple input schemas with project parameters. However, descriptions are generic and lack actionable context for LLM selection. No parameter descriptions exist, LLMs cannot infer what 'project' means without explicit guidance. Output schemas are completely undocumented; agents cannot plan downstream calls or extract structured data. No error handling guidance, no pagination support despite scanning operations that could return large result sets, and no indication of what fields are returned. Tool names follow verb_noun pattern (scan_*, list_*) but descriptions do not explain WHEN to use each tool vs. alternatives (e.g., scan_networks vs. scan_subnets vs. scan_shared_vpc). No tool annotations (readOnlyHint, idempotentHint) despite all being read-only operations.

• Add detailed descriptions to all parameters. For 'project', specify: 'GCP project ID (e.g., my-project-123). Obtain via list_projects() if unknown.'
• Document output schemas for each tool. Example for scan_instances: 'Returns array of {name, status, machine_type, zone, internal_ip, external_ip, network, subnetwork, labels, tags}. Use instance name or internal_ip for downstream tools.'
• Add pagination parameters (limit, offset/cursor) to all scan_* tools. Cap default results at 20-50 items. Document in description: 'Returns up to 50 results per call. Use offset to fetch additional pages.'
• Enhance error responses with recovery guidance. Instead of {'error': 'Permission denied'}, return {'error': 'Permission denied', 'recovery': 'Verify GOOGLE_APPLICATION_CREDENTIALS has compute.instances.list permission for this project.'}
• Add tool annotations. Mark all tools with readOnlyHint=true and idempotentHint=true to signal to agents these are safe to retry without side effects.