Claude Certified Architect — Foundations

You're designing a system where multiple specialised Claude agents work in parallel on independent subtasks, then a final agent synthesises their outputs. What is this pattern called?

An agentic workflow is processing 50,000 document summaries overnight. Latency is not a concern. Which API feature gives the biggest cost saving?

An agent loop calls a web-search tool and then re-plans based on the results. After 3 failed attempts to find relevant results, what is the correct design response?

Which of the following Claude models is the best choice for a high-volume classification subtask inside an agentic pipeline where speed and cost matter most?

In an orchestrator–subagent design, a subagent is spawned to handle a long document analysis task. What is true about the subagent's context?

You need to build a self-improving code generator where one agent writes code and another agent critiques and scores it, repeating until the score exceeds a threshold. Which pattern is this?

Why should irreversible actions (like sending an email or deleting a database record) require human confirmation inside an agentic workflow?

What distinguishes a 'retryable' error from a 'fatal' error in agentic error handling?

An agent processes user queries using Claude's tool_use feature. You notice it sometimes calls the same tool multiple times with identical inputs in a single turn. What is the most likely cause?

Which stop_reason indicates that a Claude response ended because the model wants to call a tool?

What is the primary architectural advantage of the Batch API over synchronous API calls for agentic pipelines?

A team of 20 engineers uses Claude Code on a monorepo. Where should project-wide coding standards and architecture decisions be documented so Claude Code applies them automatically?

A PreToolUse hook needs to block Claude from calling a specific shell command. What exit code must the hook script return to block execution?

Which hook fires AFTER a tool has been called and completed, and cannot block execution?

What is a Claude Code 'Skill' and how is it invoked?

You want Claude Code to automatically run a linter on every file it edits. Which hook event is the right trigger?

A subagent is spawned by the main Claude Code session to analyse a large codebase. What information does the subagent automatically have access to?

Which CLAUDE.md file location applies globally across ALL projects on a developer's machine?

The PreCompact hook fires before context compaction. What is the primary use case for this hook?

You need Claude to always respond with valid JSON matching a specific schema. What is the most reliable technique?

What is 'prefilling' the assistant turn and when should you use it?

Why should untrusted user content be wrapped in XML tags (e.g. <user_input>) rather than concatenated directly into the system prompt?

You want Claude to reason carefully through a complex maths problem before giving an answer. Which technique most directly improves accuracy?

What is the difference between the system prompt and the user message in terms of reliability guarantees?

You are designing a prompt that injects a 10,000-token knowledge base into the system prompt on every request. Which optimisation should you apply first?

Extended thinking is enabled for a request. How are thinking tokens handled in terms of cost and visibility?

A few-shot prompt provides 3 input-output examples before the actual query. What is the primary benefit of few-shot prompting over zero-shot?

What are the three core primitives of the Model Context Protocol (MCP)?

You need to connect Claude to a remote MCP server hosted on a company's cloud. Which transport mechanism is appropriate?

An MCP server needs to make an LLM call to generate a response, but it doesn't have its own Anthropic API key. What MCP feature enables this?

What makes a 'good' tool description in a tool schema, compared to a poor one?

You are building an MCP tool that deletes files. What security principle must you apply?

Which tool_choice value forces Claude to call a specific named tool, regardless of whether it thinks it needs to?

In the tool-use response cycle, what must you include in the next API request after Claude returns a tool_use block?

Your RAG pipeline chunks documents into 2,048-token segments with no overlap. Users report that answers about topics that span two chunks are often incomplete. What change fixes this?

A user's conversation has grown to 180,000 tokens in a 200k-context model. You still need to add a 5,000-token retrieved document. What should you do?

You are designing prompt caching for a RAG application. The system prompt is 3,000 tokens and changes every request; the retrieved documents are 5,000 tokens and vary per query. Where should you place the cache_control breakpoint?

When injecting retrieved documents into a prompt, what is the recommended approach for structuring the context to help Claude use it accurately?

What is the correct formula to ensure you don't exceed a model's context window when sending a request?

Across a long multi-turn conversation, which summarisation strategy best preserves important early context without consuming the full context window?