Menu
Workflow patterns and gotchas for Anthropic/Claude agents. Directs to RAG for implementation.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
This skill should be used when the user asks about "visual builder", "no-code agent builder", "drag and drop", "ThinkingConfig", "extended thinking", "chain of thought", "reasoning", or needs guidance on using ADK's visual development tools or configuring advanced reasoning capabilities.
This skill should be used when the user asks about "callbacks", "lifecycle hooks", "before_model_call", "after_tool_call", "plugins", "session state", "state management", "artifacts", "file uploads", "events", "EventActions", "human-in-the-loop", "confirmation", "memory", "MemoryService", "long-term memory", "remember across sessions", "RAG", "retrieval augmented generation", "grounding", "knowledge base", "vector search", or needs guidance on customizing agent behavior, intercepting execution, managing state across turns, implementing approval workflows, or implementing persistent memory or grounding agent responses in external knowledge.
This skill should be used when the user asks about "creating a new ADK project", "initializing ADK", "setting up Google ADK", "adk create command", "ADK project structure", "YAML agent configuration", "creating an agent", "LlmAgent", "BaseAgent", "custom agent", "agent with different model", "Claude with ADK", "OpenAI with ADK", "LiteLLM", "multi-model agent", or needs guidance on bootstrapping an ADK development environment, authentication setup, choosing between Python code and YAML-based agent definitions, agent configuration, model selection, system instructions, or extending the base agent class for non-LLM logic.
This skill should be used when the user asks about "multi-agent systems", "sub-agents", "delegation", "agent routing", "orchestration", "SequentialAgent", "ParallelAgent", "LoopAgent", "agent-to-agent", "A2A protocol", "agent hierarchy", "streaming", "real-time responses", "SSE", "server-sent events", "websocket", "bidirectional", "Live API", "voice", "audio", "video", "multimodal streaming", or needs guidance on building systems with multiple specialized agents working together or implementing real-time communication patterns.
This skill should be used when the user asks about "deploying", "production", "Agent Engine", "Vertex AI", "Cloud Run", "GKE", "Kubernetes", "hosting", "scaling", "guardrails", "safety", "content filtering", "input validation", "output validation", "authentication", "OAuth", "API keys", "credentials", "security plugins", "testing agents", "evaluation", "evals", "benchmarks", "tracing", "Cloud Trace", "logging", "observability", "AgentOps", "LangSmith", "user simulation", or needs guidance on deploying ADK agents to production environments, implementing safety measures, access control, secure authentication, testing, debugging, monitoring, or evaluating ADK agent quality.
This skill should be used when the user asks about "adding a tool", "FunctionTool", "creating tools", "MCP integration", "OpenAPI tools", "built-in tools", "google_search tool", "code_execution tool", "long-running tools", "async tools", "third-party tools", "LangChain tools", "computer use", or needs guidance on extending agent capabilities with custom functions, API integrations, or external tool frameworks.
| name | Anthropic Agents |
| description | Workflow patterns and gotchas for Anthropic/Claude agents. Directs to RAG for implementation. |
| Need | Pattern | RAG Query |
|---|---|---|
| Basic tool use | Tool definitions | "claude tool definition" |
| Agentic loop | Iterative tool calling | "claude agentic loop" |
| Computer control | Computer use | "claude computer use" |
| Deep reasoning | Extended thinking | "claude extended thinking" |
| Conversation | Message history | "claude conversation history" |
Query RAG: mcp__agentic-rag__query_sdk("pattern example", sdk="anthropic", mode="build")
These are Claude-specific traps:
tool_use vs tool_result - Tool calls are tool_use, responses are tool_resulttool_use_idmax_tokens in API callsstop_reason to know if done or needs tool responseanthropic-beta headerRAG Query: mcp__agentic-rag__query_sdk("anthropic python sdk install", sdk="anthropic", mode="explain")
RAG Query: mcp__agentic-rag__query_sdk("tool input_schema definition", sdk="anthropic", mode="build")
Tools need name, description, input_schema (JSON Schema format).
RAG Query: mcp__agentic-rag__query_sdk("messages create tool_choice", sdk="anthropic", mode="build")
RAG Query: mcp__agentic-rag__query_sdk("tool_result content block", sdk="anthropic", mode="build")
Match tool_use_id exactly in your response.
RAG Query: mcp__agentic-rag__query_sdk("agentic loop stop_reason", sdk="anthropic", mode="build")
Loop until stop_reason is not tool_use.
| Symptom | Likely Cause | RAG Query |
|---|---|---|
| Tool not called | Bad schema | "tool input_schema" |
| Tool response ignored | Wrong tool_use_id | "tool_result matching" |
| Loop never ends | Not checking stop_reason | "stop_reason end_turn" |
| Rate limit | Too many requests | "anthropic rate limits" |
| Schema validation error | Wrong JSON schema format | "json schema tool" |
Special capability for GUI automation:
RAG Query: mcp__agentic-rag__query_sdk("claude computer use setup", sdk="anthropic", mode="explain")
Requirements:
For complex reasoning tasks:
RAG Query: mcp__agentic-rag__query_sdk("claude extended thinking", sdk="anthropic", mode="explain")
Query RAG when you need:
"claude streaming response""claude image input""claude pdf document""claude prompt caching""anthropic batch api"