Anthropic Exam Guide PDF: CCA-F Study Strategy

The anthropic exam guide pdf for the Claude Certified Architect, Foundations (CCA-F) exam is the single authoritative document that defines what Anthropic will test, how questions are weighted, and what a passing performance looks like. This post decodes that guide, maps it to a prioritised study plan, and tells you exactly where to focus your hours before you sit 60 scenario-based questions scored on a 100-to-1000 scale with a passing bar of 720.

As of 3 June 2026, more than 10,000 individuals have earned the CCA-F. The exam launched on 12 March 2026 at $99 per attempt. What follows is the most direct path from the exam guide to a passing score.

What does the CCA-F exam guide actually contain?

The exam guide specifies five domains, 30 task statements, and the percentage weight each domain carries. It also describes the question format: 60 scenario-based multiple-choice items, each with one correct answer and three plausible distractors. There are no trick questions in the traditional sense; instead, every distractor represents a real engineering decision that is wrong for the specific scenario described.

The guide does not publish the raw-to-scaled score conversion, so you cannot calculate an exact question count as the pass mark. On a linear read, 720 out of 1000 corresponds to roughly 41 to 42 correct answers, but Anthropic's scaled scoring means you should target consistent performance across all domains rather than banking on a precise number.

How are the five domains weighted, and which should you study first?

Study by weight, not by listed order. The exam guide orders domains 1 through 5, but the weights tell a different story about where your time pays off.

Domain 1 alone accounts for more than a quarter of the exam. If you are short on time, every hour spent on Agentic Architecture & Orchestration returns more expected points than an hour on any other domain. Domains 3 and 4 are tied at 20% each and together represent 40% of the exam, making them the second priority block. Domain 2 at 18% is close behind. Domain 5 at 15% is the smallest slice but, as we discuss below, it is disproportionately difficult.

What does Domain 1 (Agentic Architecture, 27%) actually test?

Domain 1 tests your ability to design, debug, and evaluate multi-agent systems. The exam guide's task statements cover coordinator-subagent patterns, session management, decomposition strategies, and hook-based enforcement.

Practically, this means you need to distinguish when a hub-and-spoke architecture is appropriate versus a flat peer-to-peer topology, and you need to know what happens when a coordinator fails to pass structured context to subagents. Questions in this domain frequently present a broken agentic loop and ask you to identify the root cause from four plausible candidates.

The exam consistently rewards deterministic solutions over probabilistic ones when stakes are high. If a question describes a financial transaction pipeline and asks how to enforce a compliance rule, the correct answer is almost always a programmatic hook rather than a prompt instruction. Our concept on hooks vs prompts decision framework maps this directly to the exam's enforcement scenarios.

Key sub-topics to cover in Domain 1:

1. Coordinator responsibilities and dynamic subagent selection
2. Parallel subagent spawning and attribution preservation
3. Session management: when to resume, fork, or start fresh
4. Agentic loop anti-patterns and premature termination
5. Prerequisite gate design for multi-step workflows

How do Domains 3 and 4 (Claude Code and Prompt Engineering, 20% each) differ in question style?

Domain 3 (Claude Code Configuration & Workflows) is the most tool-specific domain. Questions test the three-level configuration hierarchy, CLAUDE.md structure, version control implications, and CI/CD integration. The three-level configuration hierarchy is a foundational concept here: global user settings, project-level CLAUDE.md, and path-scoped rules interact in ways that produce surprising behaviour when misconfigured.

Domain 4 (Prompt Engineering & Structured Output) is more analytical. Questions present a prompt and an output, then ask you to diagnose why the output deviated from the schema or how to improve reliability. The exam guide emphasises few-shot examples as the highest-leverage technique for ambiguous edge cases, and it tests your ability to choose between XML delimiters, JSON schemas, and plain prose depending on the output type required.

A useful heuristic: Domain 3 questions tend to have one clearly correct configuration answer; Domain 4 questions often require you to rank two plausible prompt strategies and choose the one that is more robust at scale. Our Prompt Engineering & Structured Output concept library covers the schema design and few-shot patterns that appear most frequently.

What makes Domain 2 (Tool Design & MCP Integration, 18%) harder than its weight suggests?

Domain 2 tests a dense surface area. The Model Context Protocol introduces its own scoping hierarchy, error flag conventions, and resource patterns that are distinct from the Messages API. Candidates who have used Claude via the API but have not worked with MCP servers often underperform here relative to the domain's 18% weight.

The exam guide's task statements for Domain 2 include:

• Writing tool descriptions that route correctly under ambiguous user intent
• Diagnosing tool misrouting caused by overlapping descriptions
• Applying the isError flag pattern for structured error propagation
• Choosing between building a custom MCP server and using an existing one
• Configuring environment variable expansion in MCP config files

The tool descriptions as selection mechanism concept is worth studying carefully. The exam presents scenarios where two tools have similar names and asks which description change would resolve misrouting with the least intervention. The answer is almost always a targeted description edit rather than a system prompt addition, because description edits are lower-effort and higher-leverage.

Why is Domain 5 (Context Management & Reliability, 15%) the trickiest domain?

Domain 5 carries the smallest weight but tests concepts that have no obvious analogue in traditional software engineering: token budget management, context rot, summary injection, and RAG retrieval strategy. Candidates with cloud certification backgrounds often find this domain the least intuitive.

The core difficulty is that context management failures are silent. A system that has lost critical context mid-session will still produce plausible-looking output, making the failure hard to detect without explicit instrumentation. The exam tests whether you can identify the conditions under which context degradation occurs and choose the correct remediation from options that all sound reasonable.

Three patterns the exam guide emphasises for Domain 5:

1. Summary injection for fresh sessions: when a session exceeds a token threshold, summarise and restart rather than continuing with a diluted context window.
2. Stale context detection: use structured fields in the context to track recency and flag when retrieved facts may be outdated.
3. RAG retrieval calibration: distinguish between retrieval failures (no relevant chunks returned) and valid empty results (the corpus genuinely contains no answer).

Our Context Management & Reliability concept library covers all three patterns with worked examples.

What scenario types appear most often in CCA-F questions?

The exam guide describes six recurring scenario archetypes. Recognising the archetype quickly is itself a time-saving skill, because each archetype has a predictable set of correct and incorrect answer patterns.

For each archetype, the exam presents a broken or suboptimal implementation and asks you to identify the root cause or the best fix. The distractor options are designed to be plausible: they represent real engineering choices that would be correct in a different scenario. The skill the exam is testing is contextual judgement, not pattern matching.

How should you structure a four-week study plan?

A four-week plan that mirrors the domain weights looks like this:

Week 4 is where a practice platform matters most. Our practice exams at AI Skill Certs are 60 questions scored 100 to 1000 with 720 as the passing bar, matching the real exam's format exactly. The adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, so it will surface your weakest task statements automatically rather than asking you to guess where your gaps are. We are independent of Anthropic; we do not claim endorsement or approval.

What hands-on projects cover the most exam ground simultaneously?

The exam guide recommends building projects that span multiple domains rather than isolated exercises. Three projects cover the most ground:

Project 1: Multi-tool agent with escalation logic Build a customer support agent that uses at least three MCP-connected tools, detects when a query exceeds its authority, and hands off to a human agent via a structured payload. This covers Domain 1 (coordinator design, structured handoff), Domain 2 (MCP tool descriptions, error propagation), and Domain 5 (context state preservation across the handoff).

Project 2: Claude Code CI pipeline Configure a Claude Code project with a CLAUDE.md at the project root, path-scoped rules for test and source directories, and a CI step that runs Claude non-interactively with the -p flag. This covers Domain 3 entirely and touches Domain 4 (structured output from the CI step).

Project 3: Multi-agent research synthesiser Build a coordinator that spawns parallel subagents to retrieve information from different sources, then synthesises with attribution. This covers Domain 1 (parallel spawning, attribution loss diagnosis), Domain 4 (structured synthesis output), and Domain 5 (context management across subagents).

Each project should be built to break: introduce a deliberate fault, observe the failure mode, and fix it. The exam tests fault diagnosis as much as correct design.

How do you simulate exam conditions without external resources?

The CCA-F is delivered online-proctored or at a test centre. You cannot use external resources during the exam. Effective simulation means:

1. Close all documentation tabs before starting a timed session.
2. Use 60-question blocks with a 90-minute timer (1.5 minutes per question is a reasonable target).
3. Flag questions where you are choosing between two plausible options rather than skipping them; review flagged questions as a set after completing the block to identify pattern gaps.
4. After each practice exam, categorise wrong answers by domain and by failure type: did you misread the scenario, misremember a concept, or reason incorrectly from correct premises? Each failure type requires a different remediation.

The concepts library at /concepts covers 174 atomic concepts mapped to all five domains and 30 task statements. Use it as a reference during study, not during simulation.