Skip to content

CCAR-F Revision Notes

Condensed notes for the final two weeks. Exam facts come from registry.yaml; SDK, Claude Code, MCP and API facts are as of 2026-10-24. Each section links to the full study guide.

Exam at a glance: 60 items (multiple-choice and multiple-response), 120 minutes, 4 scenarios drawn at random from a published bank of 6, scaled score 720 on 100–1,000 to pass, USD 125, valid 12 months. Details: README, exam day, scenarios.

Domain Weight Rough items (of 60)
1 Agentic Architecture & Orchestration 27% ~16
2 Tool Design & MCP Integration 18% ~11
3 Claude Code Configuration & Workflows 20% ~12
4 Prompt Engineering & Structured Output 20% ~12
5 Context Management & Reliability 15% ~9

Item counts per domain are our estimates from the weights, not official figures.

Domain 1 — Agentic architecture (guide)

  • Loop: tool_use → run tools, append the full assistant turn and all tool_result blocks in one user message, call again; end_turn → done; max_tokens, pause_turn, refusal → handle explicitly.
  • Never end on "text present" or parsed phrases; caps (max_turns, max_budget_usd) are safety nets.
  • Hub and spoke: subagents report to the coordinator. Coordinator selects subagents per request, partitions scope, checks coverage, re-delegates for gaps.
  • Subagents start fresh (forks excepted): pass goal, scope, prior findings and output format in the brief.
  • AgentDefinition: description + prompt required; tools allow-list restricts; model, skills, mcpServers, maxTurns optional. Spawned through the Agent tool (called Task in the v1.0 guide).
  • allowed_tools / allowedTools auto-approves; it does not restrict.
  • Parallel subagents: several Agent calls in one coordinator response.
  • Enforcement: hooks or server checks for must-hold rules; prompts for guidance. Prerequisite gates keep state outside the model.
  • Hooks: PreToolUse (permissionDecision: deny|ask|allow|defer, updatedInput), PostToolUse (updatedToolOutput, additionalContext); most restrictive decision wins; matchers match tool names.
  • Hand-offs to humans: structured summary (customer, issues, evidence, actions, recommendation, expectation).
  • Decomposition: chaining for predictable steps; dynamic for open-ended work; per-file passes + integration pass.
  • Sessions: --continue, --resume <id|name>, -n to name, --fork-session; SDK resume, fork_session / forkSession, continue_conversation / continue. Sessions hold conversation, not files. Stale context → fresh session with a summary; a few changed files → resume and name them.

Domain 2 — Tools and MCP (guide)

  • Descriptions: what, when (versus neighbours), input format with example, output, limits. Rename or split overlapping tools; check the system prompt for keyword triggers.
  • MCP errors: tool execution errors → result with isError: true; protocol errors → JSON-RPC error. Clients should give tool errors to the model.
  • Your payload convention: category (transient, validation, business_rule, permission), retryable flag, human-readable and customer-safe messages. Empty result ≠ error.
  • Subagents recover locally; escalate with attempted actions and partial results.
  • Few role-specific tools per agent; constrained tools over general ones; narrow cross-role tools only for frequent needs.
  • tool_choice: auto, none, any, {"type": "tool"}. any and forced return a 400 on Opus 5.5, Sonnet 5.5, Fable 5.1, Mythos 5.1; Haiku 5.5 accepts. Replacement: auto + instruction + strict: true + check, or output_config.format.
  • Claude Code MCP scopes: local (default, ~/.claude.json per project), project (.mcp.json, shared), user (~/.claude.json, all projects). Precedence local > project > user. ${VAR} / ${VAR:-default}.
  • Resources: @server:protocol://path; catalogues reduce exploratory calls. Tool search on by default.
  • Built-ins: Glob = paths, Grep = content, Read → Edit (unique anchor) or Write (whole file), Bash = commands. Edit fails on non-unique match → more context, replace-all, or Read + Write.

Domain 3 — Claude Code (guide)

  • CLAUDE.md: managed → user (~/.claude/CLAUDE.md) → project (./CLAUDE.md or ./.claude/CLAUDE.md) → local (CLAUDE.local.md); ancestors load at launch and concatenate; subdirectories load on demand.
  • @path imports (relative to the importing file, up to four hops). /memory to see and edit; /context to see what loaded.
  • .claude/rules/*.md: no paths → always; paths: globs → only when a matching file is read or edited. Rules beat subdirectory CLAUDE.md for conventions that follow file type.
  • Skills: .claude/skills/<name>/SKILL.md (project), ~/.claude/skills/ (personal); commands merged in; skill wins over same-name command. Frontmatter: description, argument-hint, context: fork, agent, allowed-tools (pre-approves), disallowed-tools (restricts), disable-model-invocation, paths.
  • CLAUDE.md = always-on standards; skills = on-demand procedures.
  • Plan mode (Shift+Tab, --permission-mode plan, /plan): explore and plan, no source edits until approved. Use for multi-file, multi-option, architectural work. Explore subagent = read-only discovery.
  • Iteration: concrete I/O examples, tests first, interview pattern, combined message for interacting issues.
  • CI: -p, --output-format json, --json-schema → structured_output; --allowedTools; dontAsk; --max-turns; --bare for reproducibility (then pass context explicitly). Independent reviewer; feed prior findings on re-runs; give existing tests to test generation.

Domain 4 — Prompts and structured output (guide)

  • Precision from explicit report/skip criteria and severity definitions with examples, not "be conservative". Temporarily disable noisy categories.
  • Few-shot: 2–4 diverse examples, include borderline cases with reasons, show format, show null for absent data.
  • Structured outputs: output_config.format (json_schema) or messages.parse(); strict: true tools. Shape, not truth. Incompatible with citations and prefill. Check stop_reason.
  • Schema: nullable for optional data; enum + "other" + detail; "unclear"; additionalProperties: false; normalisation rules in the prompt.
  • Validation: semantic checks in code; retry with document + failed output + specific errors; do not retry when data is absent; record which rule produced each finding.
  • Batches: 50% price, ≤24 h, no latency SLA, custom_id, results in any order, 29-day retention, ≤100,000 requests; no tool loop inside a request; submission interval = deadline − 24 h (minus margin); re-submit failures only; tune on a sample first.
  • Review: independent instance over self-review; per-file + integration passes; confidence per finding, calibrated before use.

Domain 5 — Context and reliability (guide)

  • Case-facts block outside summaries; trim tool results; key findings first, headed sections; per-issue state; resend full history.
  • Escalate: explicit request (immediately), policy gap, no progress, beyond authority. Not sentiment alone, not self-reported confidence. Multiple matches → ask for another identifier.
  • Errors: failure type, what was attempted, partial results, alternatives; distinguish access failure from empty; no silent suppression; no abort-all; coverage annotations.
  • Large codebases: subagents for discovery, scratchpad files, phase summaries, /compact, manifests for crash recovery.
  • Review: accuracy by type and field; stratified sampling of high-confidence output; calibrated field confidence; route low-confidence and conflicting items first.
  • Provenance: claim-source records preserved through synthesis; conflicts annotated with sources; dates required; established vs contested; render by content type.