Free CCA-F practice questions
Five scenario-based questions — one from each domain of the Claude Certified Architect Foundations exam. Read the scenario, pick an answer, then reveal the explanation. This is exactly the style Brevyn uses.
AAO Agentic Architecture & Orchestration
1. Your agent calls a flaky external tool. When the tool returns an error, the agent calls it again — and again — never finishing. What is the most robust architectural fix?
- A. Raise the model temperature so it tries different approaches
- B. Switch to a larger model
- C. Wrap the loop in deterministic code with a max-retry/iteration limit and explicit error handling
- D. Add "please don't loop" to the system prompt
Show answer & explanation
Correct: C.
Reliability for agent loops comes from deterministic control flow around the model, not from the model itself. Bound the loop (max iterations / retries with backoff) and handle the tool error explicitly — surface it, try a fallback, or stop. A bigger model or a prompt plea (B, D) doesn't guarantee termination; raising temperature (A) makes behavior less predictable.
CCW Claude Code Configuration & Workflows
2. You want a linter to run automatically every time Claude Code edits a file in your project, so broken formatting never lands. Which mechanism is built for this?
- A. A hook (e.g. on the file-edit tool event)
- B. A slash command you remember to run
- C. A line in CLAUDE.md asking Claude to lint
- D. A separate subagent you invoke manually
Show answer & explanation
Correct: A.
Hooks are the harness mechanism that fires automatically on events (such as after a file-editing tool runs), so the harness — not the model — guarantees the linter executes every time. A CLAUDE.md note (C) is advisory and can be skipped; a slash command (B) or manual subagent (D) rely on someone remembering.
PESO Prompt Engineering & Structured Output
3. Downstream code must reliably parse Claude's answer as JSON with a fixed shape. What gives you the strongest guarantee?
- A. Ask in the prompt for "valid JSON only" and parse the text
- B. Use a tool / structured-output schema to constrain the response, then validate it
- C. Regex the JSON out of the reply
- D. Lower the temperature to 0
Show answer & explanation
Correct: B.
Schema-constrained / tool-call output makes the shape part of the contract, and validating it on receipt (with a retry on failure) closes the loop. Asking nicely (A), regex extraction (C), or temperature 0 (D) all still leave you parsing free text that can drift — they reduce, but don't remove, the failure mode.
TDMI Tool Design & MCP Integration
4. Claude keeps misusing a tool you exposed. Which change most improves tool-use accuracy?
- A. One generic
do_action(payload)tool that does everything - B. A clear, specific name; a tight, typed parameter schema; and a description that says when to use it
- C. More tools covering every edge case
- D. Returning errors as plain prose so Claude can read them
Show answer & explanation
Correct: B.
Models use tools well when the tool is legible: a specific name, a minimal typed schema, and a description that disambiguates when to call it. A catch-all tool (A) hides intent; piling on tools (C) increases selection errors; unstructured error prose (D) is harder to act on than typed, well-defined error semantics.
CMR Context Management & Reliability
5. A long-running assistant conversation is approaching the context-window limit and per-call cost is climbing. What's the best approach?
- A. Blindly truncate the oldest messages
- B. Summarize/compact older context and keep stable, reused content cache-friendly (prompt caching)
- C. Switch to a smaller model
- D. Start every reply from an empty context
Show answer & explanation
Correct: B.
Compacting older turns into a summary preserves the information that matters while freeing tokens, and keeping the stable prefix consistent lets prompt caching cut cost and latency. Blind truncation (A) or wiping context (D) loses needed state; a smaller model (C) trades away capability rather than managing context.
Liked these? Brevyn has 200+ curated questions across all five domains, full 60-question mock exams, and unlimited AI-generated practice (Pro) that targets your weak spots. Free to start.
Brevyn is an independent exam-preparation tool and is not affiliated with, endorsed by, or sponsored by Anthropic. "Claude" is a trademark of Anthropic. Sample questions are illustrative of the exam's scenario-based style.