These 20 practice questions for Claude Certified Developer - Professional are written the way the professional exam is described: a Claude system already running in production, a piece of evidence from it (an error, a usage number, a settings file, a log line), and a decision about what explains it or what fixes it. They cover all eight domains in proportion to the weights Anthropic lists, so Production Integration, Reliability, and Knowledge Grounding gets four items and each 10% domain gets two. Three are Select TWO. Every question ends with the answer and an explanation of why each option wins or loses, because at this level the wrong options are real features applied in the wrong place.
Status on November 16, 2026
Anthropic lists Developer - Professional as a preview in its Partner Academy catalog: 70 questions (60 scored) in 135 minutes, $175, English, eight weighted domains, and no exam guide or registration yet. Everything below is checked against Anthropic's Claude Platform and Claude Code documentation. The exam guide collects what Anthropic has published, and our six Early Access practice exams give you the full-length timed version.
How these questions are built, and how to score yourself
Each stem carries one fact that decides the answer. Sometimes it is a status code, sometimes a field name, sometimes a detail about which platform or permission mode the system runs on. Read for that fact before you read the options, then eliminate everything that ignores it. Select TWO items work as two separate true-or-false judgments: an item counts only when both picks are right.
Seventeen or more out of twenty means your production knowledge is in good shape and timed full-length practice is the next step. Fourteen to sixteen is close, so target the domains behind your misses. Under fourteen means the documentation behind those domains deserves another read before more questions will help.
How the 20 questions are distributed
| Domain | Exam weight | Questions in this set |
|---|---|---|
| Production Integration, Reliability, and Knowledge Grounding | 20% | 4 (Q1 to Q4) |
| Advanced Agent Architecture and Orchestration | 17% | 4 (Q5 to Q8) |
| Observability and Production Monitoring | 12% | 2 (Q9, Q10) |
| Evaluation Engineering and Quality Assurance | 12% | 2 (Q11, Q12) |
| Runtime Guardrails and Production Security | 11% | 2 (Q13, Q14) |
| Advanced Debugging and Incident Response | 10% | 2 (Q15, Q16) |
| Agent Fleet Operations for Software Delivery | 10% | 2 (Q17, Q18) |
| Cost and Latency Optimization at Scale | 8% | 2 (Q19, Q20) |
Try 15 CCDV-P practice questions free
They come from the CCDV-P 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
The 20 practice questions
Questions 1–8
A support platform connects a Jira MCP server through the Messages API MCP connector. Its toolset sets default_config to {"enabled": false} and configs to enable search_issues, get_issue and serach_comments. Claude calls the first two tools but never searches comments, and no request ever returns an error. Which two statements are accurate? (Select TWO)
The connector merges tool settings in a fixed order: per-tool configs, then the toolset's default_config, then system defaults. Setting enabled: false as the default and enabling named tools is the documented allowlist pattern. A name in configs that the server does not have produces only a backend warning, because MCP servers can change their tools at any time, so the typo fails silently. A is wrong because that default is exactly how an allowlist is built. C is wrong because disabled tools are not offered to Claude. E is wrong because every server in mcp_servers must be referenced by exactly one toolset.
A contract-review service sends whole contracts to Claude Sonnet 5.5 with max_tokens set to 64,000. A few very large requests fail at once with a 400 invalid_request_error saying "prompt is too long". Other large requests succeed but end with stop_reason: "model_context_window_exceeded". An engineer proposes lowering max_tokens to fix both. What is going on?
When the input by itself is larger than the context window, every model returns 400 "prompt is too long". On Claude 4.5 and newer models, a request whose input plus max_tokens exceeds the window is accepted, and if generation reaches the limit it stops with model_context_window_exceeded. Lowering max_tokens affects neither case; the fix is trimming or splitting the input, counting tokens before sending, or compacting long conversations. A misreads both signals. B describes a different error, 413 request_too_large. D describes earlier models, where the beta header opts in to the stop-reason behavior.
An organization runs its production traffic in the Default Workspace and wants a $5,000 monthly spend cap on it, plus a $500 cap on a new experimentation workspace. In the Console, the experimentation cap saves, but the Default Workspace offers no spend limit at all. What should the team do?
The Default Workspace cannot be renamed, archived or deleted, and spend limits cannot be set on it. Workspace limits can keep a workspace below the organization's limits but cannot raise them, and organization-wide limits always apply. Giving production its own workspace lets it carry its own spend and rate limits. A is wrong because the Rate Limits API manages rate limits. C is wrong because an organization limit covers all workspaces together. D is wrong because the Default Workspace cannot be renamed.
An agent gives Claude web fetch, a run_python client tool that prints whatever Claude's code outputs, and a crm_lookup tool whose results contain customer websites. To stop Claude from fetching URLs it wrote into its own Python output, the team sets url_sources.client_tool_results to an except filter naming run_python. Months later someone adds a scrape_helper client tool that also echoes model-written text. What happens to URLs in scrape_helper results?
Each url_sources key defaults to {"type": "all"}. An except filter removes only the tools it names, so any client tool added later counts as a source; an only filter admits just the tools it names, so later tools are left out. Names in a filter must match tools in the request's tools array, or the request fails. A URL that appeared only in excluded sources returns url_not_in_prior_context when Claude tries to fetch it. A and B misstate the defaults, and C invents a requirement to list every tool.
A team adds on-demand compaction (compact-2026-09-04) to an agent that already uses context editing to clear old tool results. Their compaction requests return 400 errors, and they also want to enable compaction only on models that support it. Which two statements are accurate? (Select TWO)
The two features can coexist in one agent, but a single request cannot carry both compaction and context_management, so the summarization request leaves context editing out and later requests can use it again. The Models API, called with the beta header, reports capabilities.compaction for each model, which avoids discovering support through failures. A is wrong because the restriction covers only the compaction request: the SDK tool runner leaves context_management out of that request and sends it again on later ones. B invents a minimum. D replaces a documented capability check with trial and error.
A Claude Code team connects 14 MCP servers. Tool search defers their tools, which keeps the context small, but Claude often runs a search step before calling the ticketing server's create_ticket tool, which it needs on nearly every task. What fits best?
alwaysLoad exempts one server from deferral: its tools load into context regardless of the tool search setting, which suits a small number of tools Claude needs on every turn. Every upfront tool costs context, so the exemption belongs only on the servers that earn it. A is wrong because turning tool search off loads every server's tools upfront. C invents a naming rule. D is wrong because scope sets where a server is configured and which projects see it; deferral is controlled by tool search and alwaysLoad.
An Agent SDK application loads a plugin named db-tools that bundles an MCP server keyed postgres. The application sets allowedTools: ["mcp__postgres__*"], yet Claude's calls to the server's query tool are not approved. What is the tool's full name?
Tools from a plugin-bundled MCP server carry both the plugin name and the server key: mcp__plugin_<plugin-name>_<server-name>__<tool-name>, with any character outside letters, digits, _ and - replaced by _. Allow rules and allowedTools must use that full form, and a wildcard after the server prefix works as usual. A, B and D describe naming schemes that do not exist.
After upgrading Claude Code under its Agent SDK service, a team sees some subagents start with the entire conversation so far in their context, which raises cost on long sessions. The application defines no fork agent. What explains this, and how can they stop it while keeping their other subagents?
A fork is a subagent that copies the conversation so far instead of starting fresh, and from Claude Code v2.1.302 Claude can start one with nothing configured by calling the Agent tool with subagent_type: "fork". An Agent(fork) deny rule blocks that type and leaves other subagents running. A is wrong because ordinary subagents still start fresh, and omitClaudeMd only leaves CLAUDE.md files out of a subagent's starting context. B invents a mode rule. C would block every subagent along with the forks.
Full CCDV-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsEight questions in, you have seen the first two domains, which carry 37% of the exam. For the full spread at exam length, the Developer - Professional practice exams give you six timed 70-question forms.
Questions 9–12
A FinOps dashboard sums analytics usage with an api_key_ids[] filter that lists every key in the organization. Its total runs well below the invoice, and the gap matches usage from the Claude Console and from services that authenticate with Workload Identity Federation tokens. How should the query include that usage?
Requests made without an API key, such as from the Claude Console or with a Workload Identity Federation token, have api_key_id: null, so a filter listing every key misses them. Adding an empty entry to api_key_ids[] includes that usage. With a federation token, the actor is the service account the token acts as, which actor_ids[] can also target. A, C and D describe attribution rules that do not exist.
A developer on a Team plan opens /usage and sees "rate limited until 3:14am" beside the plan bars, with a note reading "Showing last-known usage". They stop working, assuming Claude Code is blocked until 3:14. What does the message mean?
When the request for plan limits fails, most often because the usage endpoint is rate limited, /usage shows the last usage bars loaded on the machine within the past 60 minutes with a "Showing last-known usage" note. A message such as "rate limited until 3:14am" means Claude Code cannot fetch the limits again until then; it does not limit the work itself. Pressing r earlier returns the same message, which rules out B. A confuses the data fetch with a usage limit, and D describes a different message about spend limits.
A team adds an agent-type PostToolUse hook that runs the linter after every Write and returns ok: false with the lint errors. They expected the turn to stop for a human, as their prompt-type hook on UserPromptSubmit does when it returns ok: false. Instead Claude keeps working and fixes the files. Which statement is accurate?
Claude Code handles an agent hook's ok: false the way it handles a prompt hook with continueOnBlock: true on the same event, so on PreToolUse and PostToolUse the turn continues with the reason fed back to Claude. Agent hooks have no continueOnBlock field at all. For prompt hooks on UserPromptSubmit, ok: false ends the turn and shows the reason as a warning line. A timed-out hook would not explain Claude fixing the lint errors, and agent hooks can block. Agent hooks are experimental, so command hooks remain the recommendation for production gates.
A team's LLM grader scores support replies from 1 to 5 against a detailed rubric and returns only the score. Agreement with human raters is good on simple replies and weak on complex ones. Which change follows Anthropic's grading guidance?
Anthropic's tips for LLM-based grading are detailed rubrics, an empirical output such as a label or a 1 to 5 score, and reasoning before the verdict through a grader that has thinking on, which helps most on complex judgments. B makes results hard to assess at scale. C removes the independence a grader needs. D throws away the rubric that makes scores consistent.
Full CCDV-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsHalfway point. If the observability or evaluation items felt unfamiliar, the Developer - Professional exam domains breakdown maps each domain to the Claude features behind it.
Questions 13–16
A company uses Claude Models on Google Cloud (Anthropic managed) and must keep inference in the US. In the Claude Console, the workspace's allowed and default inference geo settings are read-only, showing all geos allowed and a default of global. What should the team do?
On Claude Models on Google Cloud (Anthropic managed), every workspace allows all inference geos and defaults to global, and the Console shows both settings as read-only. US-only inference comes from setting inference_geo: "us" per request, priced at 1.1x on Claude 4.6 and later models. A misreads a platform rule as a role problem. B confuses workspace geo, which governs data at rest and endpoint processing and is fixed to "us" on that platform anyway, with inference geo. D cannot change a setting the platform fixes.
A security reviewer asks what leaves a developer's laptop when the developer continues a local Claude Code session from a phone with Remote Control. Which statement matches Claude Code's security documentation?
With Remote Control, the web interface connects to the Claude Code process running on the local machine, so execution and file access stay local. Session traffic travels through the Anthropic API over TLS, and while connected the transcript is stored on Anthropic servers to sync the conversation across devices. B describes cloud sessions, which are a different product. C and D misstate what is stored.
A Stop hook blocks Claude from finishing until test coverage passes 80%. On a hard task, Claude replies that it cannot raise coverage further, the hook blocks again, and after several rounds of such replies with no tool calls in between, the turn ends even though the hook blocked again. Coverage is still 72%. What explains this?
Claude Code overrides a Stop hook after it blocks eight times in a row with no tool call from Claude in between, which keeps a hook from trapping a session that cannot make progress. A hook that legitimately needs more rounds can raise the cap with CLAUDE_CODE_STOP_HOOK_BLOCK_CAP, and a hook should check stop_hook_active to know when it already triggered a continuation. A describes the budget for SessionEnd hooks. B invents a session limit. C would not produce repeated blocks at all.
A service on Claude Sonnet 5.5 sends thinking: {"type": "between_tools", "display": "summarized"} to show users short progress notes without paying for up-front thinking. Every request returns a 400 error. What should change?
between_tools is Claude Sonnet 5.5's lowest thinking setting. It turns off up-front thinking and returns each progress update with its summary text, as display: "updates" would. It takes no other field: display, budget_tokens or block_binding sent with it returns a 400 error. It is accepted only at effort high or below, so A would add a second 400. C and D add fields the setting rejects.
Full CCDV-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsFour questions left. The last two domains reward knowing how Claude Code behaves when no one is watching it: in CI, in pipelines and under organization settings.
Questions 17–20
A CI step pipes a 40 MB build log into claude -p "Find the root cause of this failure", and the step exits with an error before any analysis. Smaller logs work. What should the pipeline do?
Piped stdin is capped at 10 MB; over the cap, Claude Code exits with a clear error and a non-zero status. Writing the content to a file and naming the path lets Claude read what it needs with its tools, which also keeps a 40 MB log out of the context. B and D invent features, and C is wrong because the run fails before any turn.
A nightly claude -p run starts a background subagent to scan dependencies. Usually the run waits for the subagent before exiting, but on expensive nights it exits soon after reaching its --max-budget-usd cap, and the scan results are missing. What explains this?
A claude -p run normally stays open after Claude's turn to wait for background work, up to a 10-minute cap that CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS can change. Reaching the --max-budget-usd cap is the documented exception: Claude Code stops the remaining background work instead of waiting. If the scan matters, budget for it or run it in the foreground. A ignores both the configurable cap and the budget rule, B misreads what stops, and C invents a condition.
To control cost, IT sets maxEffortLevel: "high" in managed settings. A developer still exports CLAUDE_CODE_EFFORT_LEVEL=max, one of their subagents declares effort: max, and their user settings set maxEffortLevel: "max". Which two statements are accurate? (Select TWO)
maxEffortLevel caps the effort a session can use, and any higher level runs at the cap, whether it comes from /effort, the model picker, --effort, CLAUDE_CODE_EFFORT_LEVEL, a skill's or subagent's effort frontmatter, or the model's default. When several scopes set a cap, the lowest applies, so a cap set in managed settings cannot be raised from user settings. A per-model cap in modelSettings can exempt a model within the same settings source, which still leaves other sources' caps in force. B and C are wrong because the cap covers environment variables and frontmatter like every other source, and E is wrong because a user-level "max" sets no cap at all, so the managed cap of high still applies.
An agent on Claude Opus 5.5 runs long tool loops. Its cost model assumes thinking tokens are billed once, as output, and then disappear. Usage shows cache_read_input_tokens far larger than the visible text and tool history explains. What accounts for the gap?
Thinking blocks cannot carry cache_control themselves, but they are cached alongside the rest of a previous assistant turn, and when read from the cache they count as input tokens in usage. On models that keep previous thinking blocks, such as Opus 4.5 and later Opus models, those blocks also stay in the context window. A cost model for long agent loops has to include them as cached input. B, C and D describe billing rules that do not exist.
Full CCDV-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsTwenty questions is a diagnostic. To rehearse pacing over a full 70-question form in 135 minutes, take a timed exam from the Developer - Professional practice exams.
If you missed these, read this
A miss at this level usually means a specific behavior you have not met yet, so the fastest fix is the documentation for that feature.
- Production integration (Q1 to Q4): the MCP connector, context window and workspace sections of the exam domains breakdown, then the rules collected in the Developer - Professional cheat sheet.
- Agent architecture (Q5 to Q8): compaction, tool search and subagents, also in the domains breakdown.
- Observability (Q9, Q10): the analytics API and Claude Code usage reporting in the cheat sheet.
- Evaluation (Q11, Q12): hook types and grading methods in the cheat sheet.
- Guardrails, debugging and fleet operations (Q13 to Q18): the data residency, hook and headless sections of the cheat sheet, with the 6-week study plan for the order to study them in.
- Cost (Q19, Q20): effort caps and caching in the cheat sheet.
When the reading is done, move to volume under time pressure. The Developer - Professional practice exams give you six Early Access forms of 70 questions each, with an explanation on every answer option, and the free sample questions let you try the format first.
Frequently asked questions
Key takeaways
Key Takeaways
0/6 completedNext steps
Score yourself, note the domains behind your misses, and read the matching sections in the exam domains breakdown and the cheat sheet. Then take a timed exam from the Developer - Professional practice exams, and use the 6-week study plan to schedule the rest. For everything Anthropic has published about the exam, start with the exam guide.
Sources:
- MCP connector
- Context windows
- Workspaces
- Web fetch tool
- Compaction on demand
- Claude Code MCP and Agent SDK MCP
- Agent SDK subagents
- Usage and Cost Analytics API
- Claude Code costs
- Claude Code hooks guide
- Define success and build evaluations
- Data residency
- Claude Code security
- Thinking and prompt caching
- Run Claude Code programmatically
- Claude Code settings reference