CCA-F Exam: Complete Guide to Domains, Format, and Prep

The CCA-F exam, formally the Claude Certified Architect, Foundations certification, launched on 12 March 2026 as Anthropic's first professional credential. It costs $99 per attempt, runs 60 scenario-based multiple-choice questions, and scores on a 100 to 1000 scale with a passing mark of 720. As of 3 June 2026, more than 10,000 individuals have already earned it. This guide covers every domain, the proctored format, and the preparation strategy we recommend at AI Skill Certs.

What does the CCA-F exam actually test?

The exam tests whether you can make sound architectural decisions when using Claude in production. Every question presents a realistic scenario and asks you to choose the best solution from four plausible options. Anthropic consistently rewards deterministic solutions over probabilistic ones when stakes are high, proportionate fixes over over-engineered ones, and root-cause tracing over symptomatic patching.

There are five domains and 30 task statements. The table below shows the official weighting per Anthropic's exam guide:

Domains 1, 3, and 4 together account for 67% of the exam. If your preparation time is limited, those three domains return the most marks per hour of study.

How is the exam delivered and what are the rules?

The CCA-F is delivered either online-proctored or at a test centre. It is closed-book: no notes, no documentation, no AI assistance during the sitting. That constraint matters for preparation strategy. Because you cannot look anything up, you need to internalise the reasoning patterns behind each domain rather than memorise API signatures.

Sixty scenario-based questions at a 100 to 1000 scaled score means a passing mark of 720 is roughly 41 to 42 correct answers on a linear read, though Anthropic does not publish the raw-to-scaled conversion, so we do not treat that figure as a hard target. Aim to answer every question confidently rather than to count correct answers.

Why is Domain 1 weighted at 27% and what must you know?

Agentic Architecture & Orchestration carries the highest single weight because production Claude deployments increasingly involve multi-step, multi-agent workflows where a single bad design decision cascades into reliability failures. The exam tests whether you can design, debug, and recover these systems under realistic constraints.

The sub-topics you must be fluent in include:

1. Hub-and-spoke topology versus flat multi-agent designs. The hub-and-spoke architecture pattern centralises coordination in one orchestrator that delegates to specialised subagents. The exam tests when this is the right choice and when it introduces unnecessary bottlenecks.
2. Agentic loop anti-patterns. The agentic loop anti-patterns concept covers the most common failure modes: loops that terminate prematurely, loops that never terminate, and loops that silently swallow errors. You need to recognise each from a code snippet or scenario description.
3. Session state management. Knowing when to resume versus fork versus start fresh is a recurring scenario type. Stale context in a long-running session degrades output quality; forking creates isolation but costs tokens; a fresh session with summary injection balances both.
4. Subagent spawning and context isolation. Subagent context isolation prevents one subagent's accumulated context from polluting another's reasoning. The exam tests whether you can identify when isolation is missing and what the downstream symptom looks like.
5. Coordinator responsibilities. The coordinator responsibilities concept defines what the orchestrating agent must own: task decomposition, subagent selection, error routing, and synthesis. Questions often present a coordinator that has delegated too much or too little.

Because Domain 1 is 27% of the exam, a weak performance here is very difficult to compensate for elsewhere. We recommend spending at least a third of your total preparation time on agentic patterns.

What do Domains 3 and 4 each test at 20%?

Domain 3: Claude Code Configuration & Workflows

Claude Code is Anthropic's agentic coding tool. Domain 3 tests your ability to configure it correctly for team environments, CI/CD pipelines, and multi-developer codebases. Key areas include the three-level configuration hierarchy (user, project, and enterprise settings), version control implications of checked-in configuration files, and the design of non-interactive workflows using the -p flag.

The exam frequently presents scenarios where a configuration change at one level unexpectedly overrides another. You need to predict the outcome and identify the minimal fix.

Domain 4: Prompt Engineering & Structured Output

Domain 4 tests your ability to design prompts that produce reliable, parseable output at production scale. The prompt engineering domain covers few-shot example construction, schema design for preventing hallucinated fields, and the trade-off between over-constrained and under-constrained output schemas.

A recurring scenario type: a pipeline is receiving malformed JSON from Claude intermittently. The question asks you to identify the root cause and the proportionate fix. The correct answer almost always involves tightening the schema or adding a well-chosen few-shot example rather than adding a retry loop.

{
  "type": "object",
  "properties": {
    "summary": { "type": "string" },
    "confidence": { "type": "number", "minimum": 0, "maximum": 1 }
  },
  "required": ["summary", "confidence"],
  "additionalProperties": false
}

Setting additionalProperties: false is a low-cost, high-leverage fix the exam rewards. It eliminates an entire class of schema drift without changing the prompt.

What does Domain 2 (Tool Design & MCP Integration) test at 18%?

Tool Design & MCP Integration covers how Claude selects and invokes tools, how MCP servers are configured and scoped, and how errors propagate through tool-calling pipelines. At 18%, it is the second-smallest domain by weight, but its concepts appear as supporting context in Domain 1 scenarios as well.

The most tested concepts are:

• Tool descriptions as the selection mechanism. Claude routes to tools based on their descriptions, not their names. A vague description causes misrouting; a precise one eliminates it. The tool descriptions as selection mechanism concept explains the exact wording patterns that work.
• The isError flag pattern. MCP tools signal failure via a structured isError field rather than by throwing exceptions. The exam tests whether you can distinguish a tool that returned a valid empty result from one that failed silently.
• Tool overload. Presenting Claude with too many tools degrades selection accuracy. The tool overload problem concept covers the threshold effects and the mitigation strategies: scoping, splitting, and role-specific tool sets.

# Correct: structured error response from an MCP tool
return {
    "isError": True,
    "content": [
        {
            "type": "text",
            "text": "Database connection failed: timeout after 5000ms"
        }
    ]
}

What does Domain 5 (Context Management & Reliability) test at 15%?

At 15%, Domain 5 is the smallest by weight, but it underpins every other domain. Context management failures manifest as degraded output quality in long sessions, attribution loss in synthesis pipelines, and stale reasoning in resumed agents.

The context management domain covers progressive summarisation, summary injection for fresh sessions, and the attention dilution problem (where a very long context window causes Claude to under-weight information at the edges). The exam tests your ability to recognise which context failure mode is occurring and to prescribe the right intervention.

A common distractor answer is "increase the context window." The exam consistently rewards architects who treat context as a resource to be managed, not a parameter to be enlarged.

How should you structure your preparation?

We suggest a four-phase approach:

1. Map the domains to your existing knowledge. Score yourself honestly against the five domain areas. Most engineers are stronger on Domains 3 and 4 (Claude Code and Prompt Engineering) and weaker on Domain 1 (Agentic Architecture). Allocate study time inversely to your current confidence.

2. Build the reasoning patterns, not the memorisation. The closed-book format means you need to internalise the decision logic: when to fork versus resume a session, when to use a hook versus a prompt constraint, when to split a tool versus rewrite its description. Our concept library at /concepts covers 174 atomic concepts mapped to all five domains and 30 task statements.

3. Practice with full 60-question exams under timed conditions. The real exam is 60 questions scored 100 to 1000. Our practice exams replicate that format exactly, with the same 720 passing bar. Sitting under time pressure surfaces the gaps that reading alone does not.

4. Review wrong answers at the reasoning level, not the fact level. When you miss a question, ask why the correct answer is architecturally superior, not just which option was right. The exam rewards architects who can explain the trade-off, and that reasoning transfers to novel scenarios.

What production scenarios appear most frequently?

The exam draws scenarios from six production archetypes: customer support agents, CI/CD automation, multi-agent research pipelines, data extraction workflows, code review systems, and compliance-gated approval flows. You cannot predict which four of the six will appear in your sitting, but the underlying architectural decisions recur across all of them.

The decisions that appear most often, regardless of scenario type, are:

• Deterministic enforcement (hooks, schemas, programmatic gates) versus probabilistic enforcement (prompt instructions) when stakes are high
• Proportionate fixes (tighten a tool description, add one few-shot example) versus over-engineered fixes (add a new subagent, redesign the pipeline)
• Root-cause tracing (identify the actual failure point) versus symptomatic patching (add a retry loop around the symptom)

Mastering these three decision patterns will serve you across every scenario type the exam can present.

Is AI Skill Certs affiliated with Anthropic?

No. AI Skill Certs is an independent adaptive preparation platform. We are not affiliated with, endorsed by, or approved by Anthropic. Our platform uses Bayesian Knowledge Tracing with a 0.90 mastery threshold to identify and close your weakest domain gaps before exam day. Archie, our Socratic tutor, guides you through the reasoning behind each concept without giving answers directly, which builds the transferable pattern recognition the closed-book exam requires.

The CCA-F is part of Anthropic's Claude Partner Network, a $100M programme. As of 3 June 2026, more than 40,000 partner applicant firms have joined the network. Passing the exam is the fastest way to demonstrate production-grade Claude architecture skills within that ecosystem.