Home › CCAR-F › Cheat sheet

CCAR-F cheat sheet · 6 min read · updated 2026-09-28

CCAR-F Cheat Sheet: every fact, flag and file path the Architect Foundations exam tests

One-page revision sheet for CCAR-F: exam facts, the six scenarios, and the exact names per domain - stop_reason, allowedTools, .mcp.json, -p, tool_choice and more.

An independent community resource - not affiliated with or endorsed by Anthropic. Compiled from the official CCAR-F exam guide (Version 1.0, effective July 2026) and Anthropic's documentation.

Exam facts

The six scenarios

1. Customer support agent: Agent SDK; get_customer, lookup_order, process_refund, escalate_to_human; 80% first-contact target.
2. Code generation with Claude Code: slash commands, CLAUDE.md, plan mode.
3. Multi-agent research: coordinator plus search, analysis, synthesis, report subagents.
4. Developer productivity: exploring codebases with Read, Write, Bash, Grep, Glob and MCP.
5. Claude Code in CI/CD: automated review, test generation, PR feedback, few false positives.
6. Structured data extraction: JSON schemas, edge cases, downstream systems.

Domain 1: Agentic Architecture and Orchestration (27%)

• Loop: send, check stop_reason, run tools, append results, repeat. "tool_use" continue, "end_turn" stop.
• Anti-patterns: stopping on reply text, iteration caps as the main stop, "any text means done".
• Coordinator is the hub; subagents talk only to it. Subagents start with empty context.
• Subagents are spawned with the Task tool; "Task" must be in allowedTools.
• Parallel subagents = several Task calls in one response.
• Pass context explicitly, complete, structured, metadata separate from content.
• Prompt subagents with goals and quality criteria, not step lists.
• Narrow decomposition = coordinator failure, even when subagents succeed.
• Guarantees: hooks and prerequisite gates. Probability: prompts.
• PostToolUse hook normalises tool results; pre-call hook blocks policy violations (refund over $500) and redirects to escalation.
• Handoff to a human: customer id, root cause, amount, recommended action.
• Chaining (fixed steps) for predictable reviews: per-file passes, then integration pass. Adaptive decomposition for open-ended work: map, prioritise, plan, revise.
• --resume <name> continues a session; fork_session branches from a shared baseline; tell a resumed session which files changed; stale results mean a new session with a structured summary.

Domain 2: Tool Design and MCP Integration (18%)

• Descriptions are how the model chooses tools. Include inputs and formats, example queries, edge cases, boundaries versus sibling tools.
• Overlap: rename by purpose, split generic tools into specific ones.
• System prompt wording can override descriptions.
• Errors: isError: true; errorCategory transient / validation / business / permission; isRetryable; readable message; customer-friendly text for business rules.
• Empty result is success; failed access is failure. Keep them distinguishable.
• Fewer tools per agent (4 or 5, not 18); scope to role; constrained tools (load_document not fetch_url); scoped cross-role tool (verify_fact) for the common case.
• tool_choice: "auto" text or tool; "any" some tool; {"type":"tool","name":"x"} that tool.
• .mcp.json project scope (shared); ~/.claude.json user scope (personal); ${GITHUB_TOKEN} expansion keeps secrets out of git; all servers' tools available at once.
• Community servers for standard integrations; custom for team-specific work.
• MCP resources for catalogs (issues, docs, schemas); tools for actions.
• Grep contents, Glob paths (**/*.test.tsx), Edit needs unique anchor text, fallback Read then Write. Explore from entry points; list exported names before tracing wrappers.

Domain 3: Claude Code Configuration and Workflows (20%)

• ~/.claude/CLAUDE.md user (not shared); CLAUDE.md or .claude/CLAUDE.md project (shared); subdirectory CLAUDE.md directory.
• @import for modular standards; .claude/rules/ topic files; /memory shows what loaded.
• .claude/commands/ shared commands; ~/.claude/commands/ personal.
• .claude/skills/<name>/SKILL.md: context: fork isolates output; allowed-tools restricts; argument-hint prompts for arguments. Personal variant: different name in ~/.claude/skills/.
• Skills on demand; CLAUDE.md always loaded.
• .claude/rules/ with paths: ["**/*.test.tsx"] loads only for matching files; beats directory CLAUDE.md for scattered files.
• Plan mode: architectural, multi-file, several valid approaches. Direct: single file, clear stack trace. Plan then execute for migrations. Explore subagent for verbose discovery.
• Refinement: 2 to 3 input/output examples; tests first; interview pattern; interacting fixes in one message, independent ones in sequence.
• CI: -p / --print non-interactive; --output-format json; --json-schema. Standards and fixtures in CLAUDE.md. Prior findings for re-reviews; existing tests for test generation. Review in a separate session. CLAUDE_HEADLESS, --batch: do not exist.

Domain 4: Prompt Engineering and Structured Output (20%)

• Explicit categorical criteria beat "be conservative" and confidence thresholds. Say what to report and what to skip. Disable noisy categories to protect trust. Severity levels with examples.
• Few-shot: 2 to 4 examples for ambiguous cases, showing reasoning and format; cut false positives; handle varied document layouts; reduce false nulls.
• Tool use + JSON schema guarantees syntax, not semantics (totals, wrong fields).
• tool_choice: "any" for unknown document type; forced tool to run a step first.
• Optional (nullable) fields stop fabrication; enums with "unclear" and "other" + detail; normalisation rules in the prompt.
• Retry with document + failed output + specific error. Retry cannot find absent information.
• calculated_total vs stated_total; conflict_detected; detected_pattern on findings.
• Batches API: 50% cheaper, up to 24 h, no latency SLA, no tool-calling loops, custom_id. Overnight and weekly jobs yes; pre-merge checks no. Resubmit only failures (chunk oversized docs). Refine on a sample first.
• Independent review instance beats self-review and extended thinking. Per-file passes plus integration pass. Confidence self-reports for routing, not filtering.

Domain 5: Context Management and Reliability (15%)

• Case-facts block outside the summarised history; trim tool output to relevant fields; key findings first, section headers; full history each request; metadata with every fact.
• Escalate on: explicit request (immediately), policy gap or exception, no progress. Not on sentiment or self-reported confidence. Frustrated but resolvable: acknowledge, offer, escalate only on repetition. Multiple matches: ask for another identifier.
• Subagent errors: failure type, attempted query, partial results, alternatives. Never empty-as- success; never terminate the whole workflow. Coverage annotations in the report.
• Codebase exploration: scratchpad files; subagents for specific questions; summarise between phases; state exports and a manifest for crash recovery; /compact.
• Human review: validate by document type and field; stratified sampling of high-confidence output; field-level confidence calibrated on labelled data; route low-confidence and conflicting cases.
• Provenance: claim-source mappings preserved through synthesis; annotate conflicts with attribution; require dates; established vs contested sections; tables for numbers, prose for news, lists for technical findings.

Ruled out (do not study for this exam)

Fine-tuning; API billing and account admin; deep language or framework detail; hosting MCP servers; model internals and training; Constitutional AI and RLHF; embeddings and vector databases; computer use; vision; streaming implementation; rate limits and pricing maths; OAuth and key rotation; cloud-provider setup; benchmarks; prompt-caching internals; tokenisation.