Claude Architect Foundations

Orient the learner to the CCAR-F exam structure, domain weights, and scenario-based format.

• 60 items, 4 of 6 scenarios, 120 minutes, scaled pass 720
• Five domains: Agentic Architecture (27%), Claude Code (20%), Prompt Engineering (20%), Tool Design & MCP (18%), Context & Reliability (15%)
• Six real-world scenarios: support agent, code generation, multi-agent research, dev productivity, CI/CD, structured extraction

Establish the canonical request → tool execution → stop_reason loop that underpins every Claude agent.

• Send request → inspect stop_reason ('tool_use' vs 'end_turn') → execute tool → append result → repeat
• Tool results are appended to conversation history so Claude reasons over new evidence
• Model-driven decisions vs pre-configured trees: let Claude choose the next tool from context

Let learners step through iterations and observe how stop_reason drives termination, not assistant text.

• Observe the loop continues while stop_reason is 'tool_use'
• Predict when the loop should terminate based on stop_reason, not parsed text
• Compare correct vs anti-pattern termination logic

Explain hub-and-spoke multi-agent design with isolated subagent contexts and dynamic delegation.

• Coordinator owns decomposition, delegation, aggregation, and error routing
• Subagents do not inherit coordinator context — pass findings explicitly in the prompt
• Avoid over-narrow decomposition; design adaptive investigation plans, not fixed pipelines

Cover the mechanics of spawning subagents, allowedTools, and parallel Task calls in one turn.

• Task tool requires allowedTools to include 'Task' for the coordinator
• AgentDefinition: description, system prompt, restricted tool set
• Emit multiple Task calls in a single coordinator response to run subagents in parallel

Validate understanding of agentic loop and multi-agent orchestration before moving to tool design.

• stop_reason-based termination
• Subagent context isolation
• Coordinator's role in delegation

Distinguish programmatic enforcement from prompt-based guidance and show structured handoff patterns.

• Deterministic compliance needs hooks/pre-requisites, not just prompt instructions
• PostToolUse hooks normalize heterogeneous data formats before the model reasons
• Structured handoff (customer ID, root cause, recommended action) for human escalation

Teach how tool descriptions drive selection and how to avoid overlap and misuse.

• Tool descriptions are the primary selection signal — be specific about inputs, outputs, boundaries
• Rename and split generic tools to remove functional overlap (e.g., extract_web_results, extract_data_points)
• Audit system prompts for keyword-sensitive phrasing that overrides tool descriptions

Cover structured error metadata and per-agent tool scoping with tool_choice configuration.

• Return isError + errorCategory + isRetryable + description so the agent can decide
• Scope tools per role (4–5 tools) to keep selection reliable; replace generic tools with constrained ones
• tool_choice: 'auto', 'any', or forced {'type':'tool','name':'...'} for deterministic first steps

Walk through CLAUDE.md, Agent Skills, plan mode, and custom slash commands for team workflows.

• CLAUDE.md sets team-wide instructions read at session start
• Plan mode requires user approval before edits — use for non-trivial changes
• Custom slash commands and Agent Skills codify recurring workflows

Show reliable patterns for JSON schemas, few-shot examples, and extraction from unstructured docs.

• Provide a JSON schema and instruct the model to conform strictly
• Use few-shot examples to lock in format and edge-case behavior
• Combine schema + extraction patterns + validation hooks for high-accuracy pipelines

Synthesize context window, session resumption, and reliability patterns into a unified mental model.

• Pass context explicitly between agents; use structured formats separating content from metadata
• Choose --resume for valid prior context; start fresh with a summary when tool results are stale
• Design self-evaluation checkpoints and human-in-the-loop escalation for production reliability

Learners design the architecture, tool set, prompts, and escalation policy for Scenario 1 from the blueprint.

• Choose stop_reason loop pattern and subagent delegation strategy
• Specify MCP tool descriptions, scoping, and structured error responses
• Decide where hooks enforce compliance vs where prompts suffice
• Design handoff summary structure for human escalation

Mixed-domain assessment covering tool