How to pass the Claude Certified Architect (CCAF) exam

Begin with Anthropic's own material

Read Anthropic's exam guide first. If anything on this page differs from it, the exam guide is right.

Then take Anthropic's free courses at anthropic.skilljar.com. Three of them line up well with the exam:

• Claude Code in Action, for the Claude Code configuration and workflows domain.
• Introduction to Model Context Protocol, for tool design and MCP integration.
• Building with the Claude API, for prompt engineering and structured output.

Each has a prep quiz on our practice page, a quick check that the course material stuck.

What to study in each domain

These are the decisions to be able to make and defend in each domain, drawn from the exam guide's learning targets.

Agentic architecture and orchestration (27%)

• When a single agent is enough, and when a coordinator with sub-agents is worth the extra moving parts.
• How a long run picks up after a failure: what each step records, and how the run continues from there rather than from the start.
• What a coordinator hands to a sub-agent, and what should come back.

Tool design and MCP integration (18%)

• Designing purpose-specific tools, each with one clear job, instead of one tool that does everything.
• Returning structured errors from a tool, so the model can tell a failed call from an empty result and decide what to do next.
• Choosing between an MCP tool, a skill and a line in the prompt for a given job.

Claude Code configuration and workflows (20%)

• How the CLAUDE.md hierarchy decides which instructions apply where, and how to keep edits scoped to what a task needs.
• When to use plan mode before changes, and when a hook fits better than an instruction.
• Running Claude Code in continuous integration without leaking secrets into logs, commits or the model's context.

Prompt engineering and structured output (20%)

• Defining an output schema, validating against it, and retrying when the output does not match.
• Keeping instructions apart from untrusted content, such as a customer's message or a document being extracted, so text inside it cannot redirect the model.

Context management and reliability (15%)

• Working within a context budget: what goes into the window and what stays out.
• Writing honest failure messages, and knowing when an agent should hand over to a human.

Study each domain inside a scenario, not in the abstract. Agreeing that output needs a schema and a retry is easy. Deciding where that check belongs in a Structured data extraction pipeline that also reads untrusted documents is what the exam asks.

How to use the 120 minutes

120 minutes for 60 questions is 2 minutes each. Treat that as an average, not a limit for every question.

• Read the scenario properly. Every question sits inside one, and the answer depends on its details.
• Move on, then come back. If a question is still unclear after about two minutes, choose your best answer and move on. If the exam screen lets you mark questions for review, mark it and return once everything else is answered.
• Leave nothing blank. A best guess can score. A blank cannot.
• Check your pace at the halfway mark. At 60 minutes, aim to be around question 30. If you are well behind, give hard questions less time on the first pass.
• Judge multiple-response options one at a time. Ask of each option whether it is right for this scenario, rather than ranking it against the others.

How to use the practice modes

Our three modes suit different stages, and all of them start from the exam page.

• Learn mode checks each answer as you go and explains it. Use it while you are still learning a domain, and read the explanation even when you got the answer right.
• A practice set is untimed and graded when you submit. Use it to test yourself without help between questions.
• The timed mock runs all 60 questions against the clock. Save it for later, to rehearse pacing.

Plan around two limits. First, our bank has 80 questions, so repeat sittings show questions you have already seen. In our data, later sittings had a median score of 91.7%, against 75% for first sittings. Much of that rise can come from remembering answers, so a repeat score overstates readiness.

Second, our mock covers four of the six scenarios: the two Claude Code ones are not in it yet. Cover them through Claude Code in Action, its prep quiz and your own work in Claude Code.

When you are ready

Our data cannot tell you. Our mock is not the real exam, its 72% pass line is not Anthropic's scaled 720, and its scores do not predict the real result.

What we suggest instead is a check you can run yourself:

• You consistently pass first sittings of new material, such as a prep quiz you have not taken, rather than repeats of questions you have seen.
• You can say why each wrong option is wrong, not only which option is right.
• You have practised the two Claude Code scenarios outside our mock.

Try it

Start with one domain in Learn mode from the exam page, or look at the sample questions first: they show the answers and need no account. To gauge the difficulty first, read our guide on how hard the exam is.

CyberSkill Practice is free and unofficial, and it is not affiliated with Anthropic.