Exam Guide — Blueprint, Logistics, Strategy
The certification family (July 2026 lineup)
The Claude Certified family has four exams. You are most likely here for one of the two Architect rows. Still, it helps to see how they compare with the other two:
| Exam | Code | Audience | Q / Format | Fee | Key domain |
|---|---|---|---|---|---|
| Claude Certified Associate, Foundations | CCAO-F | Business users (ops, marketing, PM) | 60, traditional | $99 | Output evaluation (21%) |
| Claude Certified Developer, Foundations | CCDV-F | API/SDK builders | 53, traditional | $125 | Applications & Integration (33%) |
| Claude Certified Architect, Foundations | CCAR-F | Solution architects | 60, scenario-based | $125 | Agentic architecture (27%) |
| Claude Certified Architect, Professional | CCAR-P | Senior architects | 63, multiple-choice + multiple-response | $175 | Integration (19%) — Professional track → |
All four exams share these rules. Each exam lasts 120 minutes. The pass mark is 720 / 1000. Pearson VUE↗ delivers the exam. You take it online with a proctor (a remote human or AI monitor watches you through your webcam), or at a test center. The credential is valid for 12 months. If you renew on time, the renewal is free (a short refresher without a proctor, on the Partner Academy↗).
Format note: traditional means standalone multiple-choice questions; each question is independent. Scenario-based means the questions are grouped under one shared system description (see below). "ops, marketing, PM" means operations, marketing, and product-management staff. "API/SDK builders" means developers who use the Claude API (the Messages endpoint) or its software development kit (SDK).
We checked these logistics in August 2026 against two sources: the official CCAR-F exam guide (you can download it from the certification page) and the certification FAQ↗. Fees, question counts, and policies change. So download the current guide again before you pay. Some details come from the partner-tenant FAQ; they are marked, and they can change without a public notice.
CCAR-F blueprint (memorize the weights)
Before exam day, memorize these five weights. They are the most useful numbers on this page, because they tell you where to spend your study hours. The table lists the largest weight first:
| # | Domain | Weight | This guide |
|---|---|---|---|
| 1 | Agentic Architecture and Orchestration | 27% | Domain 1 |
| 2 | Claude Code Configuration and Workflows | 20% | Domain 2 |
| 3 | Prompt Engineering and Structured Output | 20% | Domain 3 |
| 4 | Tool Design and MCP Integration | 18% | Domain 4 |
| 5 | Context Management and Reliability | 15% | Domain 5 |
Officially out of scope — read, don't drill
The official CCAR-F exam guide (v1.0, July 2026) ends its content outline with a list of topics that will not appear on the exam. Many people ignore this page of the guide. But in your final review, it tells you what to skip. This KB still teaches some of these topics, because an architect needs to know them well. Each such section carries a Not on CCAR-F marker. When you see it, read the section once and move on.
| Out-of-scope topic | Still assumed | Why it's excluded |
|---|---|---|
| Cloud provider configurations (AWS Bedrock, GCP Vertex, Azure) | Claude is also served through Bedrock and Vertex. The request body is the same; the base URL and the cloud IAM are different | This is platform setup, not architecture. The Skilljar Bedrock/Google Cloud courses are optional. CCAR-P: Pro 3 · Integration |
| Fine-tuning / training custom models | On the exam, you change behavior only through prompts, tools, and context | Customers cannot fine-tune Claude |
| API auth, billing, account management | An API key exists, and it is a secret | This is operations work, not design |
| Deploying / hosting MCP servers (infra, networking, containers) | The client–server–transport model, and the choice between stdio and HTTP | Domain 4 tests tool design, not hosting. CCAR-P: integration mechanisms in Pro 3 |
| Claude internals, training, model weights · Constitutional AI / RLHF | Nothing. Treat the model as a black box with a context window | An architect cannot see these from the outside |
| Embedding models / vector DB implementation | That RAG retrieves context, and where that context enters the prompt | Deeper retrieval topics belong to CCAR-P: Pro 3 · Integration (RAG pipeline, chunking, indexing) |
| Computer use · Vision / image analysis | Nothing for CCAR-F | These are input/output features, not agentic architecture |
| Streaming / server-sent events | That streaming exists, and why a UI wants it | This is a transport detail |
| Rate limits, quotas, pricing math | Cost is a design constraint. You reason about it; you never calculate it | The numbers change every month. CCAR-P: cost/latency trade-offs in Pro 4 |
| OAuth, API-key rotation, auth protocols | Least privilege, and "don't put secrets in prompts" | This is security implementation, not agent design. CCAR-P: authn vs authz gaps in Pro 3 |
| Prompt caching implementation (beyond knowing it exists) | Put the stable prefix first; the cache lowers cost | The mechanics change. The exam tests only the idea |
| Token counting algorithms / tokenization | Context is limited, and tool output usually uses the most of it | These are vendor internals |
| Performance benchmarking, model comparison metrics | Choose a model by the type of task (speed vs depth), not by a leaderboard | These numbers go out of date fast. CCAR-P: evaluation metrics and A/B in Pro 4 |
🧭 Final-review rule: sometimes a practice question seems to depend on one of these topics. Then you are misreading it. Look for the architecture decision underneath: escalation, tool boundary, context hygiene, or deterministic step vs model step.
The scenario format — what makes CCAR-F unique
This exam is different from the others in one way. Your 60 questions do not stand alone. They come in groups under 4 scenario contexts, picked at random from a bank of 6. Each scenario is a one-paragraph description of a system. Its group of questions tests how you would design, configure, debug, and harden that one system. The good news: the six possible contexts are known in advance (full walkthroughs in Scenarios):
- Customer support agent on the Agent SDK with MCP tools and human escalation
- Claude Code for team code generation — slash commands,
CLAUDE.md, shared conventions - Multi-agent research system — a coordinator delegating to specialized subagents
- Developer productivity tooling — Read, Write, Edit, Bash, Grep, Glob + MCP server tools. These are runtime capabilities, not a list of operating-system commands. Origins, JSON arguments, shell flags, and examples.
- Claude Code in CI/CD — automated review and PR feedback
- Structured data extraction — validating output against JSON schemas
🎯 One scenario serves several questions. So if you misread a scenario, you lose a group of points, not one point. Spend a full minute on each scenario paragraph before you answer its first question. Find four things: where the model's part ends, which tools exist, which steps are deterministic (fixed code), and where the system can fail.
Registration (via the Anthropic Partner Academy↗)
⚠️ Eligibility — check this before any other step (verified from the official FAQ, August 2026): right now, certification is open only to people at Claude Partner Network organizations. To register, you need a company-domain email that is on your organization's partner record. Personal addresses (gmail, etc.) will not work. This is the reason the SSO shows "this account does not exist." If you are not part of a partner company, here is the route. Your company applies to the CPN (claude.com/form/cpn-partner-application↗). After approval, you certify under the recognized company domain. (Adding or changing a domain on a partner record takes 7–10 business days, through partner-support@anthropic.com.) The same FAQ gives two pieces of good news. First, once you register for an exam, the registration stays valid for 5 years. Second, the prep material (the courses and the downloadable exam guide) is open to everyone today. One more point clears up a common confusion. Most "Anthropic certified" posts on LinkedIn show the free Academy course completion certificates (Claude 101, AI Fluency, etc.). Anyone can earn these; Anthropic issues them officially, and you can add them to LinkedIn. They are a different thing from these proctored exams, which only partners can take. Also note what the CPN application page showed in August 2026. It gives qualification guidance for the services network: a registered business with 10 or more employees that offers consulting, implementation, or managed services. It sends product builders to the API/"Powered by Claude" track instead. Anthropic decides whether to approve an application. Being at the Registered tier is not the same as qualifying for higher tiers.
🪤 The two-tenant trap (author's experience, August 2026): there are TWO Skilljar sites, and they look the same. The public Academy (
anthropic.skilljar.com) hosts the free courses. It accepts normal email+password accounts. The certification lives on the Partner Academy tenant (anthropic-partners.skilljar.com). That site lets you sign in only through Anthropic's partner SSO. Suppose the SSO has no partner identity for your email. Then the "we'll send a code" screen sends nothing and shows no error. Google sign-in says "this account does not exist". (This is anti-enumeration: the site does not reveal whether an account exists.) In my testing, I found no way to create that identity myself. Anthropic gives access through your organization's partner record. You may believe your organization IS a partner and still fail to sign in. In that case, first check that you use your company-domain address on the recognized domain. Then contact partner-support@anthropic.com (the official FAQ says domain changes take 7–10 business days). In the end, your exam path uses four account systems: public Academy (courses) → partner SSO (exam registration) → checkout → Pearson VUE (exam delivery). Pearson VUE identifies you by the Candidate ID it emails you, and your Pearson name must match your government ID exactly.
The path from nothing to a booked exam slot is short and clear. Here are the steps, in order:
- Your organization joins the Claude Partner Network↗. New applicants start at the Registered tier. Joining opens the Partner Academy↗ to the organization's people.
- Open your certification's page on the partner tenant↗ → download the official exam guide. Do this before you pay; the guide is the official blueprint.
- Check out and pay ($125 / $175). Partner discounts apply automatically.
- Create your Pearson VUE↗ account from the link in the confirmation email.
- Book the exam online with a proctor (Pearson OnVUE) or at a test center. Both exam guides list both options. The Pearson "Select exam options" screen offers OnVUE for CCAR-F and CCAR-P (verified September 2026). Before an OnVUE exam, you must pass a system test. You can cancel or reschedule up to 24 hours before the exam.
Retakes: after a first fail, wait 14 days. After a second fail, wait 30 days. After a third fail, wait 90 days. You get at most four attempts in any 12-month period. You pay the full fee for each attempt.
Renewal: if you renew on time, the renewal is a free test without a proctor. It covers what changed. Anthropic keeps the validity period short (12 months) on purpose, because the platform changes fast.
Official preparation: Anthropic Academy↗
Anthropic Academy↗ (launched March 2026) offers free courses that you take at your own pace. These courses are the official prep. This table shows how they match the architect track:
| Course | Feeds domains |
|---|---|
| Claude Code 101 · Claude Code in Action | 2 (Claude Code), 1 (agentic workflows) |
| Introduction to MCP · MCP: Advanced Topics | 4 (Tools & MCP) |
| Building with the Claude API | 3 (prompting/structured output), 5 (context) |
| Cloud-specific courses (Amazon Bedrock = AWS · Vertex AI = Google Cloud — the cloud platforms that also serve Claude) | Not required. The CCAR-F guide lists cloud provider configuration as out of scope, and the CCAR-P guide does not name Bedrock or Vertex at all. Take these courses only if you deploy on those platforms |
| Claude 101 · AI Fluency (the rest of the shared "Foundations Prep" shelf) | The Associate (CCAO-F) track. These cover general Claude use, not architect material |
| CCAR-P prep shelf — Platform & Solution Design · Enterprise Integration & Production · Responsible AI, Safety & Risk · Stakeholder Engagement, Lifecycle & GTM · Team Enablement (≈12 h total) | One course per CCAR-P domain, across the seven domains. The Professional guide says "no single required course"; you earn the credential by passing the exam alone. See Professional |
A 3-day compressed study plan (exam in ~72 hours)
Short on time? This is the emergency version. Study the domains with the largest weights first. Do not skip the step where you redraw diagrams from memory. Drawing them again by hand is what makes you remember them.
| Day | Focus | Output |
|---|---|---|
| 1 | Domains 1 + 4 (45% combined): agent loop, orchestration patterns, tool schemas, MCP architecture | Redraw the agent-loop and MCP diagrams from memory |
| 2 | Domains 2 + 3 (40%): Claude Code config surface (CLAUDE.md, skills, hooks, settings, headless), prompting + structured output patterns |
Write a CLAUDE.md + one skill from scratch; do Samples |
| 3 | Domain 5 + Scenarios + Practice + Cram Sheet | Answer 30 practice questions; review the ones you missed; sleep |
Exam-day tactics
- 120 min / 60 Q = 2 min per question. But you read each scenario once for its whole group of questions. So plan 1 min per scenario paragraph and about 90s per question. That leaves 10 minutes for a final review.
- If a question is hard, flag it and move on. Questions in the same scenario group often make earlier questions clearer.
- Watch for the "most appropriate" stem (the question text). Several answers work. The exam wants the one that matches Anthropic's published guidance: the simplest thing that works, workflows before agents, and deterministic code where possible.
- Online proctoring needs: wired internet, a clean desk, a government ID, no second monitor, and a room scan with your webcam. A test center removes these risks. Choose it if you are unsure about your setup.
🧭 Before any scenario question, apply the five-question actor test to the question text. See the scenarios page.
🪤 The most common type of trap on architect exams: an answer is technically possible, but it breaks a constraint stated in the scenario (cost limit, latency budget, no human in the loop, compliance boundary). Read the constraint sentence again before you choose.