Claude Architect Certification Practice Questions

Domain 1 — Agentic Architecture PQ 1 / 15

Your agentic loop occasionally continues running after Claude has finished its task. The code checks if len(response.content) > 0 and response.content[0].type == "text" to determine completion. What is wrong with this approach?

A The code should check for response.content[0].type == "tool_result" instead. B The code should count the number of iterations and stop after a fixed maximum. C Claude can return text alongside tool_use blocks in the same response; checking content type does not reliably indicate completion. The code should check stop_reason instead. D The code should parse the text content for phrases like "I'm done" or "task complete."

Explanation

Text presence does not mean the agent is finished — Claude often returns explanatory text before tool calls. The stop_reason field ("end_turn" vs "tool_use") is the only deterministic signal for whether the agent intends to continue working.

Domain 1 — Agentic Architecture PQ 2 / 15

In a multi-agent research system, the synthesis subagent produces a report missing key findings that the web search subagent discovered. Logs show the coordinator invoked the synthesis agent after the web search completed. What is the most likely cause?

A The synthesis agent's system prompt does not include instructions to incorporate all findings. B The coordinator did not explicitly pass the web search results in the synthesis subagent's prompt — subagents do not inherit the coordinator's conversation history. C The web search agent stored its results in a shared database that the synthesis agent could not access. D The synthesis agent's tool set does not include a tool for reading previous agent outputs.

Explanation

Subagent context isolation means the synthesis agent starts with only what the coordinator provides in its prompt. There is no automatic inheritance of prior results. The coordinator must explicitly pass the web search findings as part of the synthesis subagent's input.

Domain 1 — Agentic Architecture PQ 3 / 15

A customer support agent needs to verify customer identity before processing any refunds. Currently, the system prompt says "Always verify the customer before processing refunds." Production logs show the agent skips verification in 8% of cases. What should you do?

A Implement a programmatic hook (PreToolUse) that blocks the process_refund tool unless get_customer has already been called and returned a verified ID. B Add more emphatic language to the system prompt: "You MUST ALWAYS verify first. NEVER skip this step." C Add 10 few-shot examples showing the correct verification-first sequence. D Reduce the agent's temperature to make it more deterministic.

Explanation

When business rules require guaranteed compliance, programmatic enforcement (hooks) is the only approach with zero failure rate. Prompt-based approaches are probabilistic — no matter how emphatic the wording, they cannot guarantee 100% adherence.

Domain 2 — Tool Design PQ 4 / 15

An agent has access to 15 tools and frequently selects the wrong one for user requests. What is the most effective architectural change?

A Write longer, more detailed descriptions for all 15 tools. B Add few-shot examples showing correct tool selection for common queries. C Implement a pre-processing classifier that selects the tool before the agent runs. D Reduce the agent's tool set to 4–5 core tools and distribute the rest across specialized subagents.

Explanation

Tool selection reliability degrades significantly above 5 tools. The architectural solution is distributing tools across specialized subagents, not writing better descriptions for 15 tools or adding a non-LLM classifier that loses context awareness.

Domain 2 — Tool Design PQ 5 / 15

Your MCP tool returns {"status": "error", "message": "Operation failed"} when it cannot connect to the database. The agent keeps retrying indefinitely. What is wrong with the error response?

A The error message should include the database connection string for debugging. B The error should be thrown as an exception rather than returned as a result. C The response lacks structured metadata (isError, errorCategory, isRetryable). Without knowing whether the error is transient or permanent, the agent cannot make intelligent recovery decisions. D The tool should silently return an empty result set instead of an error.

Explanation

Structured error responses with errorCategory and isRetryable flags let the agent distinguish transient failures (worth retrying) from permanent ones (escalate or try an alternative approach). Without this metadata, the agent defaults to retrying blindly.

Domain 2 — Tool Design PQ 6 / 15

You need to configure an MCP server that requires an API key for your project team. Where should the API key go?

A Directly in the .mcp.json file as a string value. B In an environment variable, referenced in .mcp.json using ${API_KEY} expansion syntax. C In the project-level CLAUDE.md file. D In each developer's ~/.claude/CLAUDE.md with instructions to set the key.

Explanation

Environment variable expansion (${VAR}) in .mcp.json keeps secrets out of version-controlled files. Hardcoding keys in config files is a critical security anti-pattern, and storing them in CLAUDE.md exposes them in plaintext.

Domain 3 — Claude Code PQ 7 / 15

A new team member joins and reports that Claude Code is not following the team's coding standards. Other team members have no issues. What is the most likely configuration problem?

A The team's standards are defined in a skill that the new member hasn't invoked yet. B The team's standards are in ~/.claude/CLAUDE.md (user-level) on existing members' machines instead of the project-level .claude/CLAUDE.md that would be shared via version control. C The new member has not installed the required MCP servers. D The new member needs to run /memory to load the team's configuration files.

Explanation

User-level CLAUDE.md only exists on each individual's machine. Team standards must be in project-level config (.claude/CLAUDE.md or .claude/rules/) to be shared via version control and automatically available to all team members.

Domain 3 — Claude Code PQ 8 / 15

You want to create a codebase analysis skill that produces verbose output and explores multiple files. How should you configure it to avoid polluting the main conversation context?

A Create a skill in .claude/skills/ with context: fork in the SKILL.md frontmatter. B Create a slash command in .claude/commands/ with a note to clean up after execution. C Put the analysis instructions in a subdirectory CLAUDE.md file. D Use plan mode for the analysis to keep it separate from implementation.

Explanation

context: fork runs the skill in an isolated sub-agent context. Its exploration output stays separate from the main session. Commands run in the current session and would pollute context with all the verbose output.

Domain 3 — Claude Code PQ 9 / 15

Your CI pipeline runs claude "Review this PR for security issues" and it hangs. What is the correct fix?

A Set the environment variable CLAUDE_NONINTERACTIVE=1. B Pipe the prompt via stdin: echo "Review..." | claude. C Use the -p flag: claude -p "Review this PR for security issues". D Add --headless to the command.

Explanation

The -p (or --print) flag is the documented way to run Claude Code in non-interactive mode for automated pipelines. It outputs the result directly to stdout without launching the interactive REPL. The other options reference non-existent features.

Domain 4 — Prompt Engineering PQ 10 / 15

You are building a data extraction pipeline. Claude outputs JSON that matches your schema structurally but contains fabricated values for fields where the source document had no information. What schema design change would address this?

A Add stricter validation rules that reject any null values. B Make those fields optional (nullable) in the schema so the model can return null instead of being forced to fabricate values for required fields. C Add a post-processing step that removes suspicious-looking values. D Increase the number of few-shot examples to show correct extraction.

Explanation

When source documents may not contain certain information, marking those schema fields as optional/nullable prevents the model from fabricating values to satisfy required field constraints. The model can legitimately return null instead of inventing data.

Domain 4 — Prompt Engineering PQ 11 / 15

Your extraction pipeline retries failed validations by sending "There were errors in your extraction. Please try again." The retry rarely fixes the issues. What should you change?

A Increase the retry count from 3 to 10. B Switch to a more capable model for retries. C Skip retries and route all failures to human review. D Append the specific validation errors to the retry prompt (which field failed, what was wrong, expected vs actual value) so the model knows exactly what to fix.

Explanation

Generic retry messages provide no signal for correction. Specific error feedback (field name, violation type, expected format) guides the model toward the correct output. This is the same principle as giving a human colleague specific, actionable feedback rather than "try again."

Domain 4 — Prompt Engineering PQ 12 / 15

Your team uses the Message Batches API for a weekly compliance report (processed overnight). A new requirement adds a blocking pre-commit hook that must analyze code before developers can commit. How should you handle the new workflow?

A Use the synchronous API for the pre-commit hook and continue using the Batches API for the weekly report. B Add the pre-commit analysis to the existing batch job and increase submission frequency. C Use the Batches API for both since it saves 50% on costs. D Switch everything to synchronous API for consistency.

Explanation

The Batch API processes requests within up to 24 hours, making it unsuitable for blocking workflows like pre-commit hooks that need immediate results. Use the synchronous API for latency-sensitive tasks and keep the Batch API for latency-tolerant overnight jobs where the 50% cost savings matter.

Domain 5 — Context & Reliability PQ 13 / 15

A customer support agent handles a multi-issue conversation. After 15 turns, the agent references the wrong order number when processing a refund. What context management approach would prevent this?

A Extract critical transactional facts (customer ID, order numbers, amounts, statuses) into an immutable "case facts" block positioned at the start of each prompt, outside of summarized history. B Increase the model's max_tokens to accommodate longer conversations. C Summarize the conversation every 5 turns to keep context manageable. D Add instructions to the system prompt saying "Always double-check order numbers before processing."

Explanation

Progressive summarization loses critical details. An immutable "case facts" block preserves exact values (IDs, amounts, dates) in a high-recall position (beginning of context), ensuring they are never lost or altered during conversation summarization.

Domain 5 — Context & Reliability PQ 14 / 15

Your agent's overall extraction accuracy is 96%. After deploying to production, clients report poor results on invoice documents. Investigation shows invoice accuracy is 72% while receipt accuracy is 99%. What metric design flaw allowed this?

A The validation set was too small to be statistically significant. B The test documents were not representative of production traffic. C Aggregate accuracy metrics masked per-document-type failures. The solution is to track accuracy stratified by document type. D The model was overfitting to the most common document type.

Explanation

Aggregate metrics can hide poor performance on specific segments. 99% on receipts + 72% on invoices can average to 96% overall if receipts dominate the dataset. Stratified metrics by document type reveal these hidden failures and prevent surprises in production.

Domain 5 — Context & Reliability PQ 15 / 15

In a multi-agent research system, the final report states "AI adoption grew 40% in 2025" without any source. The coordinator cannot determine which subagent produced this claim or where it came from. What architectural change would prevent this?

A Have the coordinator verify every claim by calling a separate fact-checking subagent. B Require all subagents to output structured claim-source mappings (claim text, source URL, document name, publication date) that downstream agents must preserve through synthesis. C Add a post-synthesis step that searches the internet to verify all statistics. D Limit the synthesis agent to only quoting exact text from source documents.

Explanation

Information provenance requires structured claim-source mappings from the start. Without them, attribution is irretrievably lost during summarization. Post-hoc verification is unreliable because the original source may no longer be determinable.

Claude Certification Practice Question Bank

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Explanation

Practice Bank Results

Ready for more practice?

Continue Exploring

Eligibility

Exam Content

Cost & Registration

Sample Questions

Practice Bank

How to Prepare

Career Outlook