Claude Certified Architect – Professional

1. Agentic Loop

Practice Question:
You ask Claude to fix a bug in a project. Claude reads the relevant files, modifies the code, runs the tests, sees that one test fails, changes the code again, and reruns the tests.

Which concept does this demonstrate?

A. Simple prompting
B. Agentic loop
C. Static code generation
D. Structured output

Answer: B — Agentic loop

Why?
The system is repeatedly going through:

Understand → Plan → Act → Observe → Repeat

It doesn't just generate an answer once; it takes actions, observes their results and continues working.

---

2. Tool Calling

Practice Question:
Claude needs information from a database that is not available in its existing context. What mechanism would allow Claude to retrieve that information?

A. Increase the temperature
B. Use a tool/API
C. Add more system instructions
D. Change the output format

Answer: B — Use a tool/API

Why?
Tools allow the model to interact with external systems and retrieve information or perform actions.

---

3. Read, Write, Bash, Grep & Glob

Practice Question:
You know the project contains a file named config.py, but you don't remember which directory contains it. Which capability is most appropriate?

Answer: Glob

Why?
Glob is used to find files based on file/path patterns.

Another one:

You need to find every occurrence of DATABASE_URL inside the project.

Answer: Grep

Easy reminder:

Glob → Files
Grep → Text

---

4. CLAUDE.md

Practice Question:
Your team wants Claude Code to always know that the project uses Python 3.12, pytest, and a particular project structure.

Where would you typically put project-specific instructions for Claude Code?

Answer: CLAUDE.md

Why?
It provides project-specific instructions and context that Claude can use while working on the project.

---

5. Context Window

Practice Question:
An agent is working on a large task involving conversation history, source files, instructions and many tool results.

Why is context management important?

A. It makes the CPU faster
B. It helps manage the information available to the model
C. It automatically fixes all bugs
D. It replaces tool calling

Answer: B

Why?
The model has a finite context capacity. Managing what information is retained and provided helps keep the task effective.

---

6. Context Management & Compaction

Practice Question:
An agent has been working for a long time and accumulated a large amount of conversation and tool output. What is the purpose of compaction?

Answer:
To reduce unnecessary context while retaining important information needed to continue the task.

Think about it:
Compaction isn't simply "delete everything old." The goal is to preserve useful information while making the context more manageable.

---

7. Sessions & Resuming Work

Practice Question:
Claude was working on a coding task yesterday. Today you resume the task, but another developer has changed several files in the repository.

Should Claude blindly continue based only on the previous conversation?

Answer: No.

Why?
Previous context can help, but the current state of the repository/system should be checked before continuing.

---

8. Subagents

Practice Question:
You have a large task:

Research the architecture → implement the code → run tests → review the implementation.

Would dividing these responsibilities among specialized agents be a reasonable approach?

Answer: Yes.

Why?
Subagents can be useful when a large task can be decomposed into specialized pieces.

---

9. Tool Selection & Tool Scoping

Practice Question:
A documentation agent only needs to read documentation. Should it automatically receive access to production deployment tools?

Answer: No.

Why?
Give agents only the tools and permissions they need. This reduces unnecessary access and supports the principle of least privilege.

---

10. MCP

Practice Question:
What problem does MCP primarily help solve?

A. Increasing monitor resolution
B. Standardizing connections between AI applications and external tools/data
C. Training an LLM from scratch
D. Compressing JSON files

Answer: B

Why?
MCP provides a standardized approach for AI applications to interact with external tools and data sources.

---

11. System Prompt vs User Prompt

Practice Question:
A security assistant has a system instruction saying:

"Follow the organization's security policies."

The user then asks:

"Review this Python application for vulnerabilities."

Which statement is correct?

Answer:
The system instruction establishes higher-level behavioral guidance, while the user prompt provides the specific task.

---

12. Prompt Engineering

Practice Question:
Which prompt is generally more useful?

Prompt A:

Review my code.

Prompt B:

Review this authentication module for security vulnerabilities. Focus on authentication bypass, insecure session handling and input validation. Return findings with severity, affected code and recommended fixes.

Answer: Prompt B

Why?
It provides:

Goal + Context + Constraints + Expected Output

The model has clearer information about what it needs to accomplish.

---

13. JSON

Practice Question:

What is the value of skills in this JSON?

{
  "name": "John",
  "skills": ["Python", "Cloud"]
}

Answer:
An array containing two strings.

Another question:

What is the type of:

"active": true

Answer:
Boolean.

---

14. JSON Schema & Structured Output

Practice Question:
Your application expects Claude to return customer information in a predictable JSON structure. Why might structured output/schema validation be useful?

Answer:
It helps ensure the output follows an expected structure so the application can reliably process it.

A simplified flow is:

Claude → Structured JSON → Validation → Application

---

15. Enum & Unexpected Values

Practice Question:
Your schema allows only:

house
apartment
condo

But your real-world dataset contains:

converted warehouse

What problem could occur?

Answer:
The value may fail validation because it isn't included in the enum.

Key lesson:
When designing schemas, consider legitimate values that may appear in real-world data instead of assuming the world fits perfectly into a fixed list.

---

16. Guardrails

Practice Question:
An AI coding agent can read source code and run tests. Your production database contains critical customer information.

Would giving the agent unrestricted database deletion permissions be a good design?

Answer: No.

Why?
The system should have appropriate guardrails, permissions and tool restrictions.

A safer design could be:

Read → Analyze → Create proposed change → Human approval → Sensitive action

Important:
Don't think of guardrails as only prompt instructions. Security controls should also exist at the permission, tool, authentication, validation and application levels.

---

17. Least Privilege & Security

Practice Question:
An agent only needs to read files and run tests. Which approach follows least privilege?

A. Give it administrator access
B. Give it only the required file and testing permissions
C. Give it production credentials
D. Give it access to every available tool

Answer: B

Why?
Least privilege means providing only the access necessary to perform the intended task.

---

18. Grounding & Verification

Practice Question:
Claude is asked to answer a question about a company's internal API.

Which approach is more reliable?

A. Ask Claude to guess based on general knowledge
B. Provide/retrieve the relevant API documentation and answer from that evidence

Answer: B

Why?
Grounding connects the response to relevant evidence.

For coding:

Modify → Test → Verify → Fix

is generally better than simply assuming the generated code works.

---

19. CI/CD & Automated Code Review

Practice Question:
Your organization wants an AI system to automatically review pull requests for potential bugs and security issues.

Which use case does this represent?

Answer:
AI-assisted CI/CD and automated code review.

Think about it:
The important part isn't just "Claude reviews code." You should also consider:

• What context does it receive?
• What tools can it access?
• What permissions does it have?
• What happens to its findings?
• Does a human review the result?

---

20. Human-in-the-Loop

Practice Question:
An agent has generated a production deployment plan. The deployment could cause significant business impact.

Should the agent automatically deploy without any human approval?

Answer:
Not necessarily. A human approval step can be appropriate for sensitive or irreversible actions.

A possible workflow is:

Agent creates plan → Human reviews → Human approves → Deployment

Key idea:
As the potential impact of an action increases, appropriate controls and human oversight become increasingly important.

---

One More Section I'd Add: 🧠 Quick Scenario Challenge

At the very end of the blog, I'd add 10 mixed scenario questions that combine multiple concepts.

For example:

You are building an AI coding agent. It needs to inspect a repository, modify files, run tests and potentially create a pull request. You don't want it to deploy directly to production.

Which concepts should you consider?

Answer:

• Agentic loop
• Tool calling
• Read/Write/Bash
• Context management
• Guardrails
• Least privilege
• Human-in-the-loop
• CI/CD

This is actually a better way to prepare for architecture-style questions, because one scenario can test several concepts at once.