CCARF Study Guide
The official blueprint weights five domains. Use this as a study map, then drill matching practice questions.
Agentic loops driven by stop_reason (tool_use vs end_turn), hub-and-spoke coordinators, Task-tool subagents with explicit context passing, parallel Task calls, programmatic prerequisites/hooks instead of “hope the prompt remembers,” and session resume/fork choices.
Tool descriptions that disambiguate similar tools, splitting vague tools into narrow ones, MCP isError patterns, project vs user MCP config (.mcp.json / ~/.claude.json), resources as catalogs, and picking Grep/Glob/Read/Write/Edit/Bash correctly.
CLAUDE.md hierarchy (user/project/directory), @import and .claude/rules with path globs, project slash commands and skills (context: fork, allowed-tools), plan mode vs direct execution, Explore subagent for noisy discovery, and CI-oriented Claude Code usage.
Few-shot examples for ambiguous cases, tool_use + JSON schemas for reliable structure, tool_choice (auto / any / forced), validation-retry with error feedback, Message Batches for latency-tolerant work, and independent multi-pass review instead of weak self-review.
Protect critical facts from progressive summarization, mitigate lost-in-the-middle, trim verbose tool results, escalate on clear criteria (not vibes), propagate structured errors across agents, and calibrate human review with confidence/stratified sampling ideas.
Fine-tuning, billing/auth deep dives, hosting MCP infrastructure, model internals, computer use, vision, embeddings/vector DB plumbing, and cloud-provider setup details are called out as out of scope in the official guide.