23 Claude Interview Questions for AI Engineers: API, Prompting & Agents

The Messages API gives an application direct control over prompts, history, tools, and its own agent loop. Claude Managed Agents provide a managed environment for longer-running and asynchronous agent work.

Choose the Messages API when you need custom orchestration, such as an existing support workflow that owns retrieval and permissions. Choose a managed agent when its hosted execution model fits the job. The decision criterion is control versus managed infrastructure, not model quality alone.

Example: A bank that must call its own audited customer-data service would normally retain the orchestration in its application through the Messages API. A document-analysis job that can run in a managed environment may favor a managed agent workflow.

Send a model ID, max_tokens, and one or more user/assistant messages. Claude returns a response containing typed content blocks, a stop_reason, and usage information.

Pseudo-code — replace <supported-model-id> and initialize client with the current official SDK setup:

const response = await client.messages.create({
  model: "<supported-model-id>",
  max_tokens: 300,
  messages: [{ role: "user", content: "Summarize this ticket in two bullets." }]
});

Read text from the response content blocks and handle non-normal stop reasons rather than assuming every request completes with a final answer.

Put stable role, safety, and response rules in the top-level system parameter. Put the user’s current request and variable input in a user message. The Claude Messages API uses a top-level system parameter rather than a system input-message role.

system: Answer from approved policy text. Ask for verification before account data.
user: What is the status of order ORD-1042?

This separation makes rules easier to audit and reuse, but authorization must still be enforced by application code.

max_tokens is the upper limit for generated output; it does not guarantee Claude will use that many tokens. stop_reason tells the application why generation ended, such as a normal turn completion, a tool request, or reaching an output limit.

For a report that ends because of max_tokens, do not silently present it as complete. Increase the budget, ask for a shorter output, or continue with an explicit follow-up that includes the required prior context.

Claude is Anthropic's family of AI models. Developers typically call a selected model through the Anthropic API and an official SDK, sending messages and receiving structured response content.

For example, an application can send a customer question to the Messages API, then render the returned text in its support interface. Choose a specific model ID and verify its supported features, limits, and pricing in the current documentation before deployment.

With a client-executed tool, your application receives the structured request, runs the code, and sends a tool_result back. A server-executed tool runs on Anthropic’s infrastructure according to its documented behavior.

The main decision criterion is execution ownership. Choose client tools for application-specific databases and actions where your service must enforce permissions. Regardless of tool type, limit what the model can request and validate consequential actions outside the prompt.

Example: get_customer_order is normally a client tool because the retailer’s service must verify the caller. A supported server-side web search tool can be appropriate for gathering public, current information.

Content blocks are typed pieces of message input or output. Depending on the request, they can represent text, images, tool requests, tool results, or other supported content types.

For example, a tool-using response can contain a tool_use block instead of a final text answer. An application should branch on the block type, execute only permitted tools, return a tool_result, and then request the final customer-facing response. This is safer and more reliable than parsing prose for actions.

Use a tool when the answer depends on current, private, or transactional information, or when the task must perform an external action. Use a normal model response for explanation, summarization, or reasoning over information already provided.

For example, “What is an order status?” can be explained directly, but “Where is order ORD-1042?” should call an authorized order lookup. A tool call is a request for your application to act; it is not evidence that the action is safe or permitted.

Structured outputs constrain Claude’s final response to a JSON schema, making syntax, required fields, and field types more reliable than prompt wording alone. They solve format compliance, not whether extracted values are factually correct.

For ticket routing, require:

{"queue":"billing|delivery|account","priority":"low|medium|high"}

Then validate business rules in code—for example, reject an unknown queue even if the JSON itself is valid. Confirm availability for the selected model and platform before relying on this feature.

Automatic caching lets the API manage the cache point for a growing conversation. Explicit caching marks the exact shared content block that should be reusable.

Use automatic caching for a straightforward multi-turn chat. Use an explicit breakpoint when a stable system prompt, document set, or tool definition block must remain cached while later messages vary. The decision criterion is convenience versus control over the cache boundary. Verify cache behavior with response usage fields rather than assuming a cache hit.

Prompt caching allows the API to reuse a matching prompt prefix instead of processing the same long instructions, documents, or examples on every request. Put stable content first and the changing user request after the cache breakpoint.

[stable support policy + routing examples + tool definitions]
<cache breakpoint>
User question: {{latest_question}}

For a support application, cache the policy handbook and send each new customer question last. Check cache-read usage metrics; a small edit in the cached prefix can prevent a hit.

A reliable long-document workflow gives Claude relevant source material, clear grounding rules, and enough context budget for the question and answer. Ask it to cite or quote the supplied evidence and to abstain when the evidence is missing.

For an HR handbook assistant:

1. Retrieve the most relevant policy sections.
2. Label each section with a stable ID.
3. Prompt: Answer only from these sections and cite their IDs.
4. Evaluate answers against known policy questions and conflicting passages.

Do not rely on a long context alone to guarantee factual grounding.

Treat instructions inside webpages, emails, documents, and tool results as untrusted data, not authority. Prompt boundaries help, but the primary defenses are least-privilege tools, input screening, confirmation for consequential actions, and server-side authorization.

If a fetched page says “export invoices,” the agent should summarize the page—not call an export tool. A robust workflow is:

1. Delimit untrusted content.
2. Screen or classify suspicious tool output.
3. Require explicit user confirmation for high-impact actions.
4. Test with adversarial documents before release.

Adaptive thinking lets supported Claude models decide whether and how much reasoning a request needs. Older extended-thinking modes use a manually configured thinking budget on models that support that behavior.

Use adaptive thinking for mixed workloads where some requests are simple and others require multi-step reasoning. The decision criterion is model-controlled adaptation versus manual budget control. Feature support and defaults differ by model, so verify the current model matrix rather than copying an older request configuration unchanged.

Example: A workflow that alternates between short account lookups and complex architecture reviews can let an adaptive-thinking model allocate more reasoning only where it is useful, then measure the resulting quality, latency, and cost.

The Claude MCP connector enables supported managed-agent workflows to connect to Model Context Protocol servers. It is useful when an agent needs standardized access to approved external capabilities rather than bespoke integration for every service.

For example, a research agent could connect to an organization’s approved knowledge-service MCP server. Still apply normal controls: verify the server, scope accessible tools and data, inspect results, and avoid treating MCP connectivity as an authorization mechanism by itself.

Data handling is feature-dependent. A standard Messages API request, prompt caching, uploaded files, server tools, and managed agent resources can have different retention and eligibility characteristics.

For example, an architecture may use inline document input for a short-lived analysis but avoid a file-based workflow if its retention policy does not meet requirements. The decision criterion is the actual feature and deployment platform, not the broad claim that “Claude data is retained” or “Claude data is never retained.” Review the current official retention table with legal and security teams.

Use JSON outputs when your application needs Claude’s final answer in a predictable data format. Use strict tool use when Claude must call a tool with arguments that must conform to the tool schema.

For example, a document extractor can return a JSON output such as { "invoice_total": 120 }. A shipping agent should use strict tool input for create_label({"address":"..."}). The decision criterion is data returned to your application versus arguments sent to an action.

Sampling parameters are model-specific API behavior, not universal Claude settings. Anthropic documents that some newer models reject non-default values for temperature, top_p, and top_k.

Before deploying a migration, test the complete request against the target model. If a model does not support a sampling setting, remove it and use explicit prompting, structured outputs, or evaluation to control behavior. Do not treat a 400 response as an invitation to silently fall back to an unreviewed model.