Claude Certified Architect (CCA-F) Exam Guide: Pass in 2026
Claude Certified Architect (CCA-F) exam guide: domains, weightings, scenarios, high-yield topics, and a 3-week prep plan. Start preparing today.
The Claude Certified Architect – Foundations (CCA-F) exam is Anthropic's certification for practitioners who design and implement production applications with Claude. Unlike vendor exams that test feature recall, this one tests tradeoff judgment: every question offers four plausible answers, and the three wrong ones are engineered to catch candidates with incomplete production experience.
This is the final lesson of our Claude Certified Architect course — everything you need to know about the exam itself: structure, domains, high-yield topics, the traps in the question design, and a preparation plan that works.
Previous: Lesson 6 — Context Management & Reliability
Claude Certification Exam Overview
| Detail | Value |
|---|---|
| Credential | Claude Certified Architect – Foundations |
| Questions | 60 multiple choice — 1 correct answer, 3 distractors |
| Time limit | 120 minutes (2 minutes per question) |
| Structure | 4 scenarios selected at random from a bank of 6 |
| Scoring | Scaled 100–1,000; 720 to pass |
| Cost | $125 USD per the June 2026 exam guide (earlier partner-launch coverage cited $99, with fee waivers for early partner employees) |
| Delivery | Online proctored or test center |
| Validity | 12 months |
| Prerequisites | None formal; 6+ months hands-on recommended |
Two structural details change how you should prepare. First, the exam platform requires an answer before you can advance — there's no flagging-and-skipping strategy; build the habit of committing. Second, the scenario structure means you'll spend the whole exam inside 4 production contexts (of 6 possible), so understanding the scenarios themselves is studying.
The Six Exam Scenarios
Every question is framed by one of these production contexts:
- Customer Support Resolution Agent — Agent SDK agent handling returns, billing disputes, and escalations through MCP tools, targeting 80%+ first-contact resolution
- Code Generation with Claude Code — team workflows, slash commands, CLAUDE.md, plan mode vs direct execution
- Multi-Agent Research System — coordinator delegating to search, analysis, synthesis, and report subagents
- Developer Productivity with Claude — codebase exploration with built-in tools (Read, Write, Bash, Grep, Glob) and MCP servers
- Claude Code for CI — automated reviews, test generation, PR feedback with low false positives
- Structured Data Extraction — JSON-schema-validated extraction from unstructured documents
You get 4 of the 6 at random — you cannot afford to skip a scenario in preparation.
Domain Breakdown and High-Yield Topics
Domain 1: Agentic Architecture & Orchestration — 27%
The biggest domain (~16 questions). High-yield: stop_reason loop control ("tool_use" continue, "end_turn" terminate — never text parsing or iteration caps); coordinator–subagent hub-and-spoke with explicit context passing (subagents inherit nothing); hooks for deterministic enforcement vs prompts for probabilistic guidance; narrow-decomposition failure diagnosis; --resume vs fork_session vs fresh-start-with-summary. Study Lesson 1 and Lesson 2.
Domain 3: Claude Code Configuration & Workflows — 20%
High-yield: CLAUDE.md hierarchy (user config is not shared — the "new teammate missing instructions" scenario); .claude/rules/ glob-scoped rules for scattered file types; skill frontmatter (context: fork, allowed-tools, argument-hint); plan mode for stated architectural complexity; CI flags -p, --output-format json, --json-schema (beware invented flags like --batch in distractors). Study Lesson 4.
Domain 4: Prompt Engineering & Structured Output — 20%
High-yield: categorical criteria over "be conservative"; few-shot examples that show why an action beat alternatives; tool_use schemas eliminating syntax but not semantic errors; nullable fields to prevent fabrication; retry works for format errors, never absent information; Batches API (50% savings, ≤24h, no SLA, no multi-turn tools, custom_id) — never for blocking checks; independent instance review over self-review. Study Lesson 5.
Domain 2: Tool Design & MCP Integration — 18%
High-yield: descriptions drive tool selection — enrich before adding routing layers; structured errors with errorCategory + isRetryable; empty results ≠ errors; role-scoped tool distribution (4–5 tools beat 18); .mcp.json (project, env-var expansion) vs ~/.claude.json (personal); Grep = content, Glob = paths. Study Lesson 3.
Domain 5: Context Management & Reliability — 15%
High-yield: case-facts blocks vs summarization loss; lost-in-the-middle mitigation; escalation on explicit requests/policy gaps — never sentiment or self-reported confidence; structured error propagation with partial results; segment-level accuracy before automating; claim–source provenance and temporal annotation. Study Lesson 6.
How the Distractors Are Built
Reviewing official sample questions reveals four recurring wrong-answer archetypes. Learn to spot them and you can eliminate options without knowing the topic perfectly:
- The prompt-hope answer. "Strengthen the system prompt" when the scenario involves money, compliance, or "must never happen." Deterministic requirements need hooks and gates, not better wording.
- The over-engineered answer. A trained classifier, a routing layer, a bigger model — when a low-effort fix (richer tool descriptions, explicit criteria) hasn't been tried. The exam rewards proportionate responses.
- The invented-feature answer.
CLAUDE_HEADLESS=true,--batch, acommandsarray in config JSON. If you've never seen the flag in the docs, it probably doesn't exist. - The wrong-problem answer. Technically valid but solves something the scenario didn't ask about — e.g., fixing tool availability when the failure is tool ordering.
Rule of thumb: the correct answer usually addresses the stated root cause with the smallest sufficient intervention — and preserves determinism where consequences are irreversible.
3-Week Preparation Plan
Week 1 — Build the agentic core. Work through Lessons 1–3. Build a two-subagent system with explicit context passing, one PostToolUse hook, and structured error responses. The exam's Domain 1 questions are near-impossible to reason about without having watched a coordinator fail.
Week 2 — Tooling and output. Lessons 4–5. Configure a real repo: project CLAUDE.md, two glob-scoped rules, one forked skill, one MCP server in .mcp.json. Build an extraction pipeline with nullable fields and a validation-retry loop.
Week 3 — Reliability and rehearsal. Lesson 6, then drill the 60 practice questions, then sit the full timed mock exam under exam conditions. Review every miss against its domain lesson. Book the real exam when you're consistently scoring 780+ — a buffer above the 720 pass mark.
Alongside this course, work through Anthropic Academy — Anthropic's free official training platform (13 courses on Skilljar) whose tracks like Building with the Claude API and the MCP courses map directly to the exam domains. Also read the official exam guide and documentation for the Agent SDK, Claude Code, and MCP. [EXTERNAL LINK: Anthropic's official Claude certification program page] [EXTERNAL LINK: Claude Code documentation at docs.claude.com]
What's Out of Scope
Don't waste study time on: fine-tuning, API billing/authentication, MCP server hosting infrastructure, Claude's internal architecture, RLHF/Constitutional AI, embeddings and vector databases, computer use, vision, streaming, rate limits, or cloud-provider specifics. The exam is about architecture judgment with the core building blocks, not the whole Anthropic ecosystem.
FAQ
What does CCA-F stand for?
CCA-F is the common abbreviation for Claude Certified Architect – Foundations, Anthropic's first certification tier. Higher-level certifications for sellers, architects, and developers are expected later in 2026.
How hard is the Claude Certified Architect exam?
Moderately hard for experienced builders, very hard for readers-only. Every question presents four plausible answers differing in judgment, not fact. With 6+ months of hands-on Claude experience and structured preparation, 2–3 weeks of study is typically enough.
What score do I need to pass the Claude certification exam?
720 on a scaled 100–1,000 score. The scaled model equates difficulty across exam forms, so the raw-question equivalent varies slightly — aim for 80%+ on practice tests for a safe margin.
How long is the Claude Certified Architect certification valid?
12 months from the date awarded — notably shorter than the 3 years typical of AWS or TOGAF credentials, reflecting how fast the Claude platform evolves. Plan for annual recertification.
Is there a prerequisite for the Claude Architect Foundations exam?
No formal prerequisite. Anthropic describes the ideal candidate as a solution architect with 6+ months of practical experience across the Claude API, Agent SDK, Claude Code, and MCP.
How much does the Claude certification cost?
$125 USD, delivered online proctored or at a test center.
Is the Claude certification worth it?
For practitioners already building with Claude professionally, yes — it is the only credential that validates hands-on skills with the Agent SDK, MCP, and Claude Code, and it launched alongside Anthropic's $100M Claude Partner Network, so partner organizations actively look for it. If you have never built anything with Claude, build first: the certification validates skill rather than substituting for it.
Is the Anthropic certification free?
The preparation is free — Anthropic Academy offers 13 free courses covering the exam domains, and this course plus our practice bank and mock exam are also free. The exam itself costs $125 per the June 2026 exam guide, with fee waivers reported for early partner employees.
Where do I start preparing?
Start with our free Claude Certified Architect course — 7 lessons mapped to the exam domains — then validate with the practice question bank and timed mock exam.
Conclusion
The Claude Certified Architect – Foundations exam certifies the skill that actually matters in production AI engineering: choosing the right pattern among several that work. Master the five domains in weight order, internalize the distractor archetypes, build the hands-on exercises, and rehearse under timed conditions.
Next step: test yourself against the 60-question practice bank, then sit the full timed mock exam. If you're new to the material, start at Lesson 1.