CCAR-F Objective Checklist (Self-Assessment)

Tick each box when you can explain it out loud and justify the trade-off, not just recognize it. Each objective links to where it’s taught in this chapter.

Domain 1 — Agentic Architecture & Orchestration (27%)

• Distinguish workflow (code-defined path) vs agent (model-defined path); default to the simplest. → AGENTIC_PATTERNS_DEEP_DIVE.md
• Match work-shape to pattern: prompt chaining, routing, orchestrator-workers, parallelization (sectioning/voting), evaluator-optimizer, dynamic decomposition. → same
• Explain orchestrator-workers vs sectioning by who decides the subtasks. → same
• Use subagents for isolated context / parallelism; know when a single agent is better. → AGENT_SDK_DEEP_DIVE.md
• State the agent loop (gather → act → verify → repeat). → same
• Pick among Agent SDK / CLI / Client SDK / Managed Agents. → same

Domain 2 — Claude Code Configuration & Workflows (20%)

• CLAUDE.md memory hierarchy (enterprise → user → project → subdirectory) and how levels combine. → CLAUDE_CODE_DEEP_DIVE.md
• CLAUDE.md (instructions) vs settings.json (behavior/permissions). → same
• Custom slash commands; MCP prompts surface as slash commands. → same
• Plan mode = research/propose, no changes until approval. → same
• Hooks enforce deterministic guardrails vs prompt guidance. → same/AGENT_SDK
• Least-privilege permissions and tool approval. → same

Domain 3 — Prompt Engineering & Structured Output (20%)

• Stateless Messages API; system prompt sent every request; attention decays with length. → CLAUDE_API_DEEP_DIVE.md
• Few-shot > prose for format; principles > conditionals for judgment; explicit conditionals for safety-critical. → same
• Tool use loop; four tool_choice modes (auto/any/tool/none). → same
• Structured outputs via JSON Schema (output_config.format) vs prompt-only JSON; schema > prose. → same
• Nullable/optional fields to prevent fabrication; semantic validation; correction-not-blind-retry. → same
• Extraction tool vs native structured output — the two mechanisms. → same

Domain 4 — Tool Design & MCP Integration (18%)

• “Make the right action easy, the wrong action hard.” Enums, stable IDs, structured errors, pagination-on-demand. → AGENTIC_PATTERNS_DEEP_DIVE.md
• Tool composition: bundle mechanical steps, keep decision points separate. → same
• MCP primitives: tools (model), resources (application), prompts (user) — and which to use. → MCP_DEEP_DIVE.md
• Annotations are untrusted hints, not security. → same
• Protocol errors vs tool execution errors (isError: true). → same
• MCP scope precedence: local > project > user. → same
• Progressive availability / list_changed / tool search. → same

Domain 5 — Context Management & Reliability (15%)

• “The model sees a request, not your database.” → CONTEXT_AND_RELIABILITY_DEEP_DIVE.md
• Context strategies: sliding window, summarization, structured state, persistent reference, retrieval, tool-result compression, native compaction. → same
• Relevance beats volume even with huge windows. → same
• Error classification (transient/validation/business/permission) and correct responses. → AGENTIC_PATTERNS_DEEP_DIVE.md
• Uncertain writes: idempotency / check-then-act, never blind retry. → CONTEXT_AND_RELIABILITY_DEEP_DIVE.md
• Confidence calibration → human review; feedback loops. → same
• Prompt caching: stable prefix first, volatile last; changes invalidate cache. → CLAUDE_API_DEEP_DIVE.md

Cross-cutting “always true” answers

• Prefer structure over prose for reliability.
• Push each concern to the right owner: model / application / tool / schema.
• Prefer the simplest pattern that meets the requirement.
• Least privilege + human approval for irreversible actions.

If every box is checked and you can defend the trade-off, you’re ready. Do a final cold pass of PRACTICE_QUESTIONS.md.