Claude Certified Architect — Foundations practice questions

Question 1

While an orchestration thoroughness mode is being built, a developer assumes it sets a hidden API parameter and a secret effort level above xhigh. Per the documented orchestration-mode design, what is actually true?

Accepted Answer

The mode is not an API parameter; it is assembled from a documented Effort value like xhigh plus a mid-conversation reminder.

Answer

The mode is a dedicated API parameter that unlocks an effort level sitting above the documented xhigh ceiling for requests.

Answer

The mode is a permission rule in settings.json that auto-approves parallel subagent fan-out on a strict per-request basis.

Answer

The mode is a hook matcher that rewrites the cached system prefix once every several turns to keep the mode active.

Question 2

An architect is about to let Claude Code refactor a payments module that spans many files. They want Claude to investigate and produce a reviewable plan that touches no source until they approve, and they understand they can edit the plan before approving. Which permission mode and approval flow fits, and what happens on approval?

Accepted Answer

Use plan mode: Claude reads files and proposes changes without editing, and approving the plan exits plan mode into the chosen editing mode.

Answer

Use acceptEdits mode so Claude edits files as it goes and the architect reviews the resulting diff afterward in a single consolidated pass.

Answer

Use a custom subagent to do the refactor in its own context, returning only a summary so the main session never makes any edits at all.

Answer

Use default mode and answer each permission prompt per tool call, since default mode already blocks all edits until the final approval step.

Question 3

Having already moved untrusted documents into tool results, a developer is taking the next recommended hardening step: making the nature and origin of that content explicit. Where does the guidance say to convey what the content is and where it came from?

Accepted Answer

In the tool's description, or in the structure of the result itself, stating its nature and source

Answer

In a regex post-processing pass that strips out any embedded instructions before the model reads it

Answer

In the response schema, by adding a required provenance field that the model must populate on each call

Answer

In the system prompt, by enumerating every trusted domain the agent is permitted to read from

Question 4

A developer is designing a Slack MCP tool for an agent that often reads many threads in one session. A detailed response (about 206 tokens) returns thread content plus channel_id, user_id, and thread_ts identifiers, while a concise response (about 72 tokens) returns only thread content. Most calls just need to read messages, but some follow-up calls require thread_ts to fetch replies. What is the most token-efficient design that still supports the follow-up calls?

Accepted Answer

Add a response_format parameter defaulting to concise, and let the agent request detailed responses only when it needs the IDs.

Answer

Return the detailed format but strip the message bodies, keeping only the identifier fields to minimize the per-call payload size.

Answer

Always return the detailed format on every call so the thread_ts identifier is guaranteed present whenever a reply fetch is needed.

Answer

Always return the concise format and have the agent reconstruct any required thread_ts from the visible message text by parsing it.

Question 5

A platform team is hardening a Customer Support Resolution Agent built on the Agent SDK over MCP tools (get_customer, lookup_order, process_refund, escalate_to_human). They want first-contact resolution above 80% but a safe handoff when the model cannot proceed. The agent loop currently only checks for stop_reason 'end_turn' and 'tool_use'. Which gap is most important to close first?

Accepted Answer

Handle stop_reason 'max_tokens' and 'refusal' explicitly so truncated or safety-stopped turns trigger continuation or escalate_to_human rather than silently failing.

Answer

Treat every stop_reason 'tool_use' as a refund approval and call process_refund immediately, since the model already decided to act.

Answer

Ignore non-'end_turn' stop reasons and simply retry the same request three times before surfacing a generic error to the customer.

Answer

Cache the system prompt and tool schemas so the agent loop runs faster and the unhandled stop reasons never get a chance to occur.

Question 6

A developer is building the Developer Productivity agent over built-in tools plus an MCP codebase server. They want Claude to search the web and run code, but they prefer Anthropic to execute those operations rather than wiring execution into their own service. Which tool category meets that preference?

Accepted Answer

Server tools such as web_search, code_execution, web_fetch, and tool_search run on Anthropic's infrastructure.

Answer

Outcome events, which run the requested shell operations as a direct part of grading the session.

Answer

Client tools, including user-defined and bash/text_editor schema tools, always run in your own app.

Answer

MCP tools, which by their definition always execute remotely on Anthropic's own servers directly.

Question 7

To run two Claude Code sessions at once — one building a feature, one fixing a bug — without their edits colliding, a team is setting up git worktrees. Which statement about worktree isolation is correct?

Accepted Answer

Each worktree is a separate working directory with its own branch but a shared history and remote, so edits in one session never touch files in another.

Answer

Each worktree is a full independent clone with its own separate repository history and remote, so a commit in one never reaches the other checkout at all.

Answer

The worktrees share one working directory but keep separate branches, so the two sessions still edit the same files and must coordinate every write.

Answer

Worktrees isolate only untracked files, so a tracked-file edit in one session will immediately overwrite that same tracked file in the other session.

Question 8

A platform team is migrating a structured-extraction pipeline to Claude Opus 4.8 and notices it no longer infers unstated steps from a single example the way the prior model did, especially at lower effort. How should they understand this behavior?

Accepted Answer

Opus 4.8 follows instructions literally and will not silently generalize or infer unmade requests

Answer

It is a regression; Opus 4.8 has dropped support for few-shot examples inside extraction pipelines

Answer

Raising the temperature sampling parameter will restore the prior model's tendency to generalize

Answer

The model needs its entire system prompt removed so that it stops over-reasoning about the task

Question 9

An analysis subagent in the Multi-Agent Research System returns large structured payloads to the coordinator. The team is unsure whether to format tool responses as JSON, XML, or Markdown to maximize the coordinator's downstream reasoning quality, and they want a defensible way to decide. What does Anthropic's tool-writing guidance recommend?

Accepted Answer

There is no one-size-fits-all format; select the response structure by running your own evaluations since the optimal choice varies by task.

Answer

Always return JSON because it is the only format whose schema the model can parse reliably across every agentic task.

Answer

Always return Markdown because models trained on web text universally reason better over Markdown than over JSON or XML.

Answer

Always return XML because its closing tags give the strongest structural signal and remove all formatting ambiguity for any task.

Question 10

A team building a Structured Data Extraction service sends the same long JSON-schema instructions and a large reference document on every request, varying only the short input record to extract from. Latency and per-request cost are too high. Which use of prompt caching gives the biggest, behavior-neutral win?

Accepted Answer

Cache the large, stable prefix (schema instructions plus reference document) so subsequent requests with the same prefix reuse it automatically.

Answer

Move the short varying input record to the front of the prompt and cache it, so the most frequently changing block benefits from caching first.

Answer

Route each extraction through the Message Batches API, which guarantees identical latency to caching while halving the live token cost.

Answer

Cache the model's output blocks after each call so the extracted JSON is replayed verbatim on the next structurally similar record.

Question 11

A team is starting a Managed Agent session and is confused about the lifecycle: they sent the agent ID and environment ID but the agent is idle. They expected work to begin on creation. What does the session lifecycle actually require, and what does passing the agent ID as a string select?

Accepted Answer

First create the session to provision its sandbox, then send a user event; the string ID uses the latest version.

Answer

Sessions auto-run on creation, and passing the agent ID as a string pins an explicit older version.

Answer

A session needs only an environment ID, and the agent ID string just names the sandbox to be reused.

Answer

Creation both provisions and starts work, and the agent ID string always references a fixed environment.

Question 12

An enterprise rolls out Claude Code fleet-wide via Intune and needs deterministic precedence when multiple managed settings drop-in files define overlapping values: telemetry settings should load before security settings, and a later security file must win for any conflicting scalar value. How should they author the drop-in directory to guarantee this merge order?

Accepted Answer

Use numeric filename prefixes such as 10-telemetry.json and 20-security.json so files sort alphabetically and later files override earlier scalars.

Answer

Rely on alphabetical-by-content sorting and trust that the security file's values override telemetry because security is the more sensitive concern.

Answer

Hide the telemetry file by prefixing its name with a dot so it is ignored, leaving only the security file to define all of the scalar values.

Answer

Place both settings inside managed-settings.json as a single base file, since the base is merged last and therefore wins all scalar conflicts.

Question 13

A team operating an agentic Claude Code workflow in CI finds the model favors reasoning over actually invoking its repo-inspection and test-running tools, leaving reviews incomplete. They are already at high effort. Which TWO levers does Anthropic's best-practices guidance recommend to increase the level of tool usage? (Select two.)

Accepted Answer

Adjust the prompt to explicitly instruct the model about when and how to properly use its tools during the workflow.

Accepted Answer

Raise the effort setting, since higher effort shows substantially more tool usage in agentic search and coding workloads.

Answer

Disable extended thinking entirely so reasoning tokens cannot crowd out the tool-invocation portion of each turn.

Answer

Set the temperature parameter above its default to inject variability that nudges the model toward calling tools more often.

Question 14

A developer is streaming Claude's responses and sees that when a tool_use block starts its input is an empty object, then a flurry of input_json_delta events arrive. Their handler JSON-parses each delta on arrival and keeps throwing parse errors. What is the correct way to obtain the tool's full input?

Accepted Answer

Concatenate the partial_json fragments from every input_json_delta and parse the assembled string once the block closes.

Answer

Parse each input_json_delta as it arrives; any fragment that fails to parse is malformed and should be dropped before the next.

Answer

Read the complete input from the content_block_start event, since its empty object is replaced in place as the deltas stream.

Answer

Issue a follow-up tools/list call after the block closes to fetch the finalized input the server persisted for that call.

Question 15

To keep a 200-page policy manual out of every prompt, a team is looking for a way to upload the document once and point each Messages request at it rather than pasting the full text on every call. Which Claude capability fits?

Accepted Answer

Upload the manual once through the Files API to get a file_id, then reference that file_id in each Messages request instead of re-uploading the content.

Answer

Paste the manual into the system prompt once and rely on automatic prompt caching to keep the whole document cached for free across every later request.

Answer

Save the manual with the memory tool so Claude writes it to disk and reloads the entire document back into context at the start of every new request.

Answer

Split the manual across several batch requests so the Message Batches API holds the text server-side and merges it into later interactive calls for you.

Claude Certified Architect — Foundations

Sample questions

Sources

Similar exams