Here are twenty free CCA-F practice questions written the way the Claude Certified Architect - Foundations exam asks them: a short production scenario, an implementation decision to make, four options that all describe real Claude features, and one option that fits this situation best. Every question carries a domain tag, an answer, and an explanation that covers why the right option wins and why each distractor loses. The set follows the published blueprint (five questions on agentic architecture, four each on Claude Code, prompt engineering, and tool design, three on context management), and five of the twenty are "Select TWO" items, close to the share you should expect on the real exam. Work through them untimed first, then use the scoring rubric to decide where to study next.
Start Here
New to the exam? Read the CCA-F complete guide for the domain map and the seven anti-patterns, keep the CCA-F cheat sheet open while you work, and when you are ready for full-length timed practice, the CCA-F practice tests on preporato.com are built to the same 60-question, five-domain format. Start with the free CCA-F sample questions (no signup) to see the question style and find your weakest domain before you buy.
How to use these 20 questions
CCA-F items are best-fit questions. Every option is usually a real feature or a real pattern; the wrong ones are wrong for the scenario in front of you, so read the stem for constraints (a shared team, a hard "never", a user waiting on the result) before you look at the options. When two options both seem defensible, the reliable tiebreaker across all five domains is that programmatic enforcement (a schema, a hook, a stop_reason check, a manifest file) beats prompt-based guidance.
Score yourself honestly, counting a "Select TWO" item as correct only when both picks are right:
- 17 or more out of 20: you are ready for timed full-length practice; book once you clear that consistently.
- 14 to 16: borderline. Read the explanations for every miss, then study the domains you missed most.
- Under 14: work through the complete guide and the topic articles linked in the closing table before more questions.
The real exam gives you 60 questions in 120 minutes, about two minutes each, so a reasonable target for this set is 40 minutes.
Build the stop_reason loop and schema pipeline the tiebreaker rewards
Programmatic enforcement is easier to spot under exam time once you have wired it yourself: one graded project has you architect an agentic loop with safe termination on stop_reason, the other a structured-output pipeline where a schema, a validation pass, and a retry that carries the error do the enforcing.
- Architect a Claude Agentic Loop with Safe Termination ~3hintermediateRubric-graded ProOpen project
- Build a Structured-Output Extraction & Review Pipeline with Claude ~3hintermediateRubric-graded ProOpen project
Try 15 CCA-F practice questions free
They come from the CCA-F practice tests, and each option shows its explanation once you answer.
- 6 timed practice exams
- Every answer explained
- A score for each domain
- $19.99 one-time
CCA-F dumps, and why they do not transfer
People searching for CCA-F exam dumps are usually after the same thing these questions give: real items, in the real format, with the reasoning shown. The difference is that a dump gives you the answer key to questions that will not be on your exam, and this set gives you the decision rule that will be.
That is not a moral point, it is a structural one about how CCA-F is built. Every item is best-fit: four options that all describe genuine Claude features, one of which fits the constraints in this stem. Change the stem from a shared team to a solo developer, or add a user waiting on the result, and the correct option changes while all four options stay the same. Memorising an option letter against a remembered question teaches you nothing that survives that swap. Candidates who prepare this way report recognising the shape of a question and still having no way to choose between two plausible options, because the thing being tested was never the recall.
There is a practical risk too. Anthropic delivers CCA-F through the Partner Academy under terms you accept when you register, and distributing or using leaked items is the kind of thing certification programmes revoke credentials over. Weigh that against a $125 retake.
What actually transfers is the tiebreaker. Across all five domains, when two options both look defensible, the one that enforces the behaviour programmatically wins: a schema over an instruction, a hook over a convention, a stop_reason check over an assumption, a manifest over a habit. That single rule resolves more CCA-F items than any amount of memorised question text, and every explanation below shows it being applied.
If you want volume of real-format practice, that is a reasonable thing to want. The free sample questions needs no signup, and the full CCA-F practice tests run 390 questions across six timed exams with every distractor explained.
Agentic Architecture & Orchestration (27%): questions 1-5
Questions 1–5
An insurance company runs a claims-intake system as a hub-and-spoke (a coordinator agent that routes work to specialized subagents). The coordinator collects the policy number and incident details from the customer, then delegates to a fraud-screening subagent. In testing, the subagent repeatedly reports "no policy number provided" even though the coordinator confirmed it two turns earlier. What is the likely cause and the correct fix?
Subagents do not inherit the parent's conversation. Context isolation is deliberate: it keeps each worker's window clean and predictable, and it means everything the worker needs must be passed explicitly in the task prompt. Option C misdiagnoses the problem, since max_tokens limits output length and has nothing to do with what a subagent sees. Option B adds tools and repeated lookups where a single passed value would do, and pushes the subagent's tool count up. Option D removes a sound separation of concerns to work around a bug that one line in the delegation prompt fixes.
A compliance team's agent produces a weekly regulatory digest. A coordinator delegates three research jobs (EU, US, and UK rule changes) to subagents, then a fourth subagent writes the digest from all three sets of findings. The current design runs the four subagents strictly one after another and takes forty minutes. Which TWO changes are correct? (Select TWO)
The three regional jobs are independent, so they belong in one coordinator turn as parallel task calls; wall-clock time drops to roughly the slowest region. The digest depends on all three results, so it stays sequential behind them, which is option C. Option A launches a dependent step before its inputs exist and produces a digest that must be rewritten. Option D breaks context isolation: subagents should receive only the scoped brief they need, and sharing the whole conversation adds cost and noise without improving coordination, because merging results is the coordinator's job.
An IT helpdesk agent resets passwords, unlocks accounts, and assigns software licenses. The operations lead wants a policy for when the agent hands a ticket to a human. Tickets sometimes stall when the agent retries the same unlock call with the same arguments and nothing changes. Which escalation trigger should be implemented programmatically?
No progress after N attempts is one of the three reliable triggers (with an explicit request for a human and an ambiguous policy), and it is observable from the tool log without asking the model anything. Option C relies on self-reported confidence, which is poorly calibrated. Option A uses sentiment, which fires on informal language and misses polite frustration. Option D confuses complexity with ambiguity: a multi-tool ticket that sits inside clear policy is exactly the work the agent should finish on its own.
A support operation runs one Claude Code investigation session per fraud case. Day-shift analysts stop mid-investigation, and the night shift must continue from the exact point where they left off, with the full history of tool calls and findings intact. Separately, a senior analyst wants to test an alternative theory on one case without disturbing the main line of investigation. Which pairing is correct?
Resuming a named session restores the conversation history exactly, which is what a shift hand-off needs. Forking creates an independent branch from the same baseline, so the alternative theory can diverge without writing back to the main line. Option B throws away tool history and uses resume for a job that needs isolation. Option A reverses the two: a fork for the hand-off creates a second branch nobody merges, and a fresh session for the theory loses the shared baseline. Option D depends on a process never dying, a fragile assumption for a multi-day case.
A logistics company's agent re-plans routes when a truck breaks down. The agentic loop (send a message, run any tool Claude requests, send the result back, repeat) calls a routing tool, a driver-notification tool, and a fleet database. In production it sometimes stops before drivers are notified, and sometimes keeps calling the routing tool after a plan is final. How should the loop decide when to continue and when to stop?
The Messages API returns a deterministic stop_reason on every response, and it is the supported signal for loop control: tool_use means run the tool and continue, end_turn means the model is done, and max_tokens means the output was cut off and needs continuation handling. Option C parses free text, which is why drivers sometimes go unnotified (the phrase appears mid-task) and the loop sometimes runs on (the phrase never appears). Option A caps on an average, so long incidents get truncated. Option B turns a self-report into control flow, the same fragility as C with an extra call. Keep an iteration or token budget as a logged safety net only.
Full CCA-F set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsStudy tip
Domain 1 rewards matching the architecture to the task and passing context on purpose. Before your next pass, read Claude Code subagents and orchestration patterns.
Tool Design & MCP Integration (18%): questions 6-9
Questions 6–9
An HR team's pipeline sends every incoming resume to Claude with three tools defined: extract_candidate, check_duplicate, and notify_recruiter. Every resume must produce a structured candidate record on the first call, yet about one response in twenty comes back as a prose summary instead of an extract_candidate call. Which change fixes this most reliably?
tool_choice controls whether and which tool Claude calls, and forcing a named tool is the programmatic guarantee: the response will be an extract_candidate call that conforms to its input schema. Option D removes prose but still lets the model pick check_duplicate or notify_recruiter on a resume that has not been extracted yet. Option A is prompt-based guidance and remains probabilistic, which is the current failure. Option C drops schema enforcement entirely and relies on prompt-only JSON, which the exam treats as the weakest structured-output approach.
A platform team is standardizing Claude Code for twelve engineers. Everyone needs the same Jira MCP server (an external tool server speaking the Model Context Protocol), which authenticates with a token that must never be committed. Several engineers also run personal MCP servers (note-taking, music) that teammates should not inherit. Which TWO configuration decisions are correct? (Select TWO)
Project-level .mcp.json is the shared, version-controlled scope, and ${VAR} expansion keeps the secret out of the file, so option D gives the whole team one definition and no leaked token. User-level configuration is the right home for servers only one person wants, which is option C. Option B makes the shared file unshareable and still leaves a plaintext secret on disk. Option A duplicates the team server twelve times and lets it drift; variable expansion already keeps the secret local while the server definition stays shared.
A retail support agent has two tools: lookup_customer, described as "Gets customer info", and find_customer, described as "Finds customer data". The first takes an exact customer ID; the second does a fuzzy search on name and email. The agent frequently calls the ID tool with a name and gets a not-found error. What should you change first?
Tool descriptions are the primary routing signal, and these two are near-identical, so the model has nothing to choose on. A description that states input format, output, and a boundary ("for partial names use find_customer") fixes selection at the source. Option C helps only marginally: names carry far less weight than descriptions. Option D moves routing logic into the system prompt, where it competes with everything else and drifts away from the tool definitions. Option B forces a tool call but does nothing to help the model pick the right one.
You are writing an MCP tool that charges a card through a payments gateway. The gateway returns HTTP 429 when the merchant is rate-limited and HTTP 403 when the API key lacks the charge permission. Today the tool returns an empty result in both cases, and the agent tells customers the payment "did not go through". How should the tool respond instead?
Structured errors let the agent decide the next action: back off and retry a transient 429, or stop and escalate a permanent 403. Both outcomes differ from an empty result, which the model reads as "the operation completed and found nothing". Option B is unstructured, so the retry decision falls to free-text interpretation. Option A hides a permission failure behind endless retries and can hang the loop. Option D keeps the silent-empty anti-pattern for the retryable case and crashes the loop on the case that should be reported and escalated cleanly.
Full CCA-F set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsClaude Code Configuration & Workflows (20%): questions 10-13
Quick check
A platform lead wrote the repository's testing conventions (which runner to use, how to name fixtures) into ~/.claude/CLAUDE.md on their laptop. CLAUDE.md is the instructions file Claude Code loads at start; teammates' sessions ignore the conventions. The lead also keeps a personal preference for terse responses that should never affect anyone else. Which TWO placements are correct? (Select TWO)
User-level CLAUDE.md is personal and never travels with the repository, which is exactly why teammates see nothing. Shared conventions belong at project level, where every clone loads them (B), and personal preferences belong at user level (C). Option D copies files by hand and drifts the moment the lead edits the original. Option A makes always-on conventions depend on someone remembering to invoke a skill, and skills are meant for on-demand procedures, while standing rules belong in CLAUDE.md.
Full CCA-F set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsStudy tip
When a stem says "must", "never", or "always", the exam is steering you toward a hook or a permission rule rather than an instruction. The mechanics are in Claude Code hooks explained.
Never write secrets to .env files.
Step the same secret-writing request through both guards: as a CLAUDE.md line the advice gets weighed and skipped and the key lands in .env, while as a PreToolUse hook the script exits 2 and the write is blocked before permissions are checked, which is the must-or-never signal in questions 10 to 13.
Questions 11–13
A fintech team has two hard requirements for Claude Code in its payments repository: the agent must never edit files under migrations/, and the black formatter must run after every edit to a Python file. A CLAUDE.md line for each already exists, and both requirements are still violated in long sessions. Which mechanism enforces them?
Hooks are shell commands the Claude Code harness runs at lifecycle events, outside the model's discretion. A PreToolUse hook that exits 2 blocks the tool call and feeds the reason back to Claude, and a PostToolUse hook runs the formatter after every real edit. Option B is still prose, loaded conditionally; it advises and cannot block. Option A improves compliance but stays probabilistic, which is the failure being reported. Option C stops the wrong edits only if the human catches them, slows every change, and does nothing to run black.
A CI pipeline invokes Claude Code on every pull request to produce a list of review findings that a later job turns into GitHub comments. Some runs hang until the job times out, and on runs that do finish, the parsing step fails on the output. Which invocation fixes both problems?
The -p (print) flag makes Claude Code non-interactive, so it never waits for input a pipeline cannot give, and --output-format json produces machine-readable output that a downstream job can parse deterministically; a schema check turns any malformed output into a clear failure. Option B keeps an interactive session alive in a headless job and parses prose. Option D fixes the hang but relies on prompt-only JSON. Option C treats a configuration problem as capacity and leaves the parsing failure in place.
A security team wants a /dependency-audit skill (a reusable, on-demand instruction set for Claude Code) that scans lockfiles, runs an audit command, and produces a long report. Two constraints: the report's volume must not fill the main conversation, and the skill must never be able to modify files. Which frontmatter satisfies both?
context: fork runs the skill in its own context, so the audit's verbose output never lands in the parent conversation, and allowed-tools is the least-privilege boundary that makes file modification impossible where an instruction would only discourage it. Option B leaves the file-writing tools available and asks nicely. Option C removes Bash (needed to run the audit) and keeps every line of output in the main window. Option A lets the edits happen and cleans up afterward, which is both riskier and noisier than preventing them.
Full CCA-F set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsPrompt Engineering & Structured Output (20%): questions 14-17
Questions 14–15
A procurement platform extracts purchase-order fields with Claude. Two defects recur: about 2% of responses are not parseable JSON, and in another few percent the line-item amounts do not add up to the stated total. The current prompt says "return valid JSON with these fields" and includes six examples. What is the right design?
Three layers map to the defects: the schema (a JSON Schema attached to a tool definition, so the response must match it) removes malformed JSON, validation code catches totals that do not reconcile, and the retry sends the specific validation error back so the correction is targeted. Option A overstates what a schema does; it enforces syntax (fields and types), never semantics such as sums. Option B keeps the prompt-only approach that is already failing and pushes examples past the 2-4 range where they help. Option D doubles cost on every order and still cannot say which of two disagreeing outputs is right.
A publisher has two workloads: re-tagging 400,000 archived articles, due within two days, and a writer's assistant that suggests tags in a multi-turn chat while an article is being drafted. Finance wants the Batch API's 50% saving applied wherever it is safe. Which TWO decisions are correct? (Select TWO)
The Batch API (asynchronous submission of many requests, processed within a 24-hour window at half price) fits the archive: it is a background job with a two-day deadline, and no user is waiting on it. The assistant fails both Batch constraints at once: someone is waiting on the result, and it needs multi-turn conversation, which the Batch API does not support. Option D blocks a live user on a job that might return in minutes or many hours. Option A pays full price for a workload that gains nothing from real-time delivery.
Full CCA-F set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsStudy tip
Domain 4 turns on the three-layer pattern: schema for syntax, validation for semantics, retry with the error attached. Work through structured output and prompt engineering for CCA-F before you retry questions 14 to 17.
Step the clean-source run to watch validation reject a string where a number belongs and a currency outside the enum, with the exact errors carried into the retry, then switch to malformed source to see the budget run out and escalate to a human: the three layers questions 14 to 17 keep testing.
Questions 16–17
A community platform uses Claude to flag posts for moderator review. The prompt says "be careful not to over-flag harmless posts", yet moderators report that most flagged items are benign, and the queue has become unmanageable. Which revision most directly reduces the false positives?
Vague guidance ("be careful") gives the model nothing measurable, so its threshold floats. Explicit criteria ("flag only if the post contains X, Y, or Z") make each decision checkable, and a few examples showing acceptable borderline content teach what should stay unflagged. Option D is more of the same vagueness, louder. Option A overshoots the 2-4 example range, consumes context, and often produces inconsistent pattern-matching. Option C spends money to compensate for an unclear specification that a bigger model will still read as unclear.
A real-estate extraction schema has a parking_type enum with garage, street, and none. Listings that say only "parking available" get labeled garage, and downstream valuations are skewed. Some listings say nothing about parking at all, and those must stay distinguishable from the ambiguous ones. How should the schema change?
Two different signals need two different values: null for absent information and unclear for information that is present but ambiguous. Giving the model a legitimate way to say "mentioned, type unknown" stops it from forcing a wrong category, and routing those rows to review keeps the valuation clean. Option B collapses both cases into null, so downstream cannot tell "no parking mentioned" from "parking mentioned, type unknown". Option A manufactures data from a plausible-sounding guess. Option D throws away the enum's consistency and hands the classification problem to whoever consumes the free text.
Full CCA-F set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsContext Management & Reliability (15%): questions 18-20
Questions 18–20
A due-diligence agent reads vendor contracts of 60 to 90 pages and answers a fixed checklist. Reviewers find that indemnity and limitation-of-liability clauses in the middle third of a contract are missed far more often than clauses near the start or the end. Which change addresses the cause?
This is the lost-in-the-middle effect: attention is strongest at the start and end of a long context, and material in between is under-weighted. Reducing the context per call (section passes, then a synthesis over the extracts) removes the middle instead of asking the model to try harder. Option B enlarges the very thing causing the problem. Option C is prompt-based guidance against a positional bias, and max_tokens governs output length only. Option A adds structure but keeps the whole contract in one call, so the middle stays the middle.
A nightly agent enriches 20,000 product records by calling Claude and two catalog APIs. Records are independent, and the run must finish before the store opens. Last week the process crashed at record 12,400, restarted from record 1, and doubled the night's spend. Which TWO changes give the job reliable crash recovery? (Select TWO)
A manifest is the durable, external record of what was done: the task list, per-item status, and any partial results. Reading it on restart is what makes recovery skip completed work instead of repeating it, and the two halves only work together. Option D is what the team effectively has today, and it guarantees the doubled cost. Option A stores state inside a context window that disappears with the crash and asks the model to track something code should track.
A claims-adjudication system issues 50,000 automated decisions a day; roughly a third are denials. Ten reviewers can check about 500 decisions daily and currently sample 1% at random. An audit found that denials above a monetary threshold carry four times the error rate of everything else. What should change?
A fixed review budget buys the most quality assurance when it is aimed where errors concentrate, and the audit has already located them. Stratified sampling (reviewing a high share of high-risk outputs and a low share of routine ones) plus error tracking per stratum turns review into a feedback loop for prompts and validation. Option D is over 16,000 items a day for a team that can do 500. Option B stays random, so most of the extra effort lands on low-risk decisions. Option C is self-reported confidence, which is poorly calibrated and skips the errors the model is confident about.
Full CCA-F set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsStudy tip
Domain 5 questions almost always hinge on one anti-pattern (bigger context, sentiment triggers, self-reported confidence, silent empty results). The full list with fixes is in common CCA-F exam mistakes to avoid.
Score yourself and pick your next read
Tally your misses by domain, then use the table. Two or more misses in a domain means that domain gets the next study block; a single miss usually means one concept, so read the explanation again and move on.
If you missed these, read this
| Domain | Questions | What the misses usually mean | Read next |
|---|---|---|---|
| Agentic Architecture & Orchestration (27%) | 1-5 | Context isolation, parallel vs sequential delegation, escalation triggers, session resume vs fork, stop_reason | Subagents and orchestration patterns; complete guide, Domain 1 |
| Tool Design & MCP Integration (18%) | 6-9 | tool_choice modes, project vs user MCP scope with env expansion, descriptions as routing, structured errors | MCP tool design best practices; cheat sheet, Domain 2 tables |
| Claude Code Configuration & Workflows (20%) | 10-13 | CLAUDE.md hierarchy, hooks vs instructions, -p and JSON output for CI, skill frontmatter | Hooks explained; CLAUDE.md context management; permissions and settings precedence |
| Prompt Engineering & Structured Output (20%) | 14-17 | Schema plus validation plus retry, Batch API constraints, explicit criteria, nullable vs unclear | Structured output and prompt engineering for CCA-F |
| Context Management & Reliability (15%) | 18-20 | Lost-in-the-middle mitigation, manifest-based recovery, stratified human review | Complete guide, Domain 5 and the seven anti-patterns |
The linked articles: subagents and orchestration patterns, MCP tool design best practices, Claude Code hooks explained, CLAUDE.md context management, permissions and settings precedence, and structured output and prompt engineering. For the numbers on where candidates actually lose points, see the hardest CCA-F topics from practice attempt data.
Build the tool set and skill behind questions 6-13
Repeated misses in Tool Design or Claude Code Configuration usually point at mechanics you have read about without wiring: tool descriptions that route correctly, project-scoped MCP config, skill frontmatter, and permission rules; one graded project has you design a tool set with an MCP server for a Claude agent, the other build a skill, a command, and a permission set.
- Design Tools and an MCP Server for a Claude Agent ~3hintermediateRubric-graded ProOpen project
- Build a Claude Code Skill, Command, and Permission Set ~3hintermediateRubric-graded ProOpen project
Frequently asked questions
Key Takeaways
0/8 completedNext steps
Once you have your domain tallies, take the free CCA-F sample questions for a second data point, then move to the full CCA-F practice tests, which are weighted to the same five domains and explain every distractor. If you need a schedule to hang the study blocks on, the 30-day CCA-F study plan maps them week by week. All six tests, the flashcard deck, and the graded tasks are included in Preporato Pro; see pricing for the current options.
Sources:
![CCA-F Practice Questions With Explanations: 20 Free Scenarios [2026]](/blog/cca-f-practice-questions-with-explanations-2026.webp)