When building Gemini/LLM integrations, always use the most current and indexed API documentation/spec first, and avoid inefficient fallback behaviors. Apply this standard: - If MCP tool `search_documentation` is available: treat it as the only documentation source. - Call `search_documentation`, read the returned docs, and then generate code immediately. - Do not manually fetch URLs. - Do not fall back to `llms.txt` after MCP already provided sufficient information. - Cap documentation lookups (e.g., maximum 2 tool calls) before generating output. - If MCP is not available: use the official docs index (`llms.txt`) to discover pages, then fetch only the specific needed pages. - For API schemas/fields/operations: use the REST API discovery spec as the source of truth (default v1beta; use v1 only when explicitly pinned). - When adding new skills (e.g., Interactions API), prevent use of deprecated SDKs and base/test your implementation against the existing `gemini-api-dev` skill examples. Example (MCP-first behavior): ```text IF tool search_documentation exists: docs = search_documentation("How to call Structured Outputs in Gemini") // Ensure docs are sufficient generate code using docs // Do NOT fetch llms.txt as a fallback ELSE: fetch llms.txt index fetch only the needed page(s) generate code using fetched docs ```