Independent guide
Start here · Step 1

Sign in with your partner-organisation email on the Skilljar site, not claude.ai

Create an Anthropic Academy (Skilljar) account, authenticate, and accept the EULAs — then every course and resource below unlocks.

Important: this is a separate account from claude.ai — not your Claude login. Sign up fresh with your partner-organisation email.

Sign in / sign up at the Anthropic Partner Academy ↗

An independent study guide · Claude Certified Architect - Professional

Pass the
CCAR-P.

Own the full lifecycle of a production Claude system. The exact prep path, the exam blueprint, and a guide to all 7 domains.

  1. Sign in at Skilljar with your partner-organisation email — not claude.ai.
  2. Practice on personal gear only — never run Claude on employer devices, networks, or data.
  3. Follow the prep path below (about 12 hrs).
  4. Book the exam — 63 Qs · 120 minutes · closed-book · pass 720/1,000.
Sign in with your partner email ↗ See the prep path →
63questions
120minutes
720to pass /1,000
7domains
01

Your prep path

This is the official Anthropic Academy prep path for this exam. Complete these courses in order — together they cover every domain. Each course opens on Skilljar — sign in first.

  1. 01 Course
    Claude Platform & Solution Design
    238 min Start ↗
  2. 02 Course
    Enterprise Integration & Production
    158 min Start ↗
  3. 03 Course
    Responsible AI, Safety & Risk for Architects
    114 min Start ↗
  4. 04 Course
    Stakeholder Engagement, Lifecycle & GTM
    178 min Start ↗
  5. 05 Course
    Team Enablement & Operational Productivity
    45 min Start ↗

That’s the whole prep. Then practice hands-on (use the domain guides below as reference) and book the exam (sign in with your partner email first). ≈ 12.2 hrs total

02

Exam blueprint

A proctored, closed-book exam: 63 scenario-based multiple-choice / multiple-response questions in 120 minutes, no AI assistance. Scored to 1,000 with a passing bar of 720. It is proctored online or at a Pearson VUE test center, and the credential is valid for 12 months from the date the credential is awarded.

CCAR-P
exam code
63
questions
$175
exam fee
120 minutes
time limit
720/1,000
passing score
12 months
validity

Domains & weights

Study time should follow the weights. Click a domain to jump to its guide.

01 Solution Design & Architecture 17%
02 Claude Models, Prompting & Context Engineering 13%
03 Integration 19%
04 Evaluation, Testing & Optimization 16%
05 Governance, Safety & Risk Management 14%
06 Stakeholder Communication & Lifecycle Management 14%
07 Developer Productivity & Operational Enablement 7%
03

The five domains

Reference for each domain — the concepts to know, the official docs, and exam tips. The prep path above is what to take; this is what to know. Click a domain to expand it.

01

Solution Design & Architecture

17%

Turning an ambiguous business problem into a defensible Claude architecture. This is the domain that decides whether everything downstream is solving the right problem.

Key concepts to master
Workflow vs. agent. A workflow follows a path you defined; an agent decides its own path. Workflows are cheaper, more predictable and easier to debug — choose an agent only when the task genuinely cannot be enumerated in advance.
The augmented LLM. A single model call enriched with retrieval, tools and memory. The simplest pattern that works, and the one candidates skip past too quickly.
Orchestrator-worker. A coordinating agent decomposes a task and delegates to focused sub-agents. Buys parallelism and separation of concerns at the cost of latency and token spend.
End-to-end architecture. Input, processing, output and the feedback loop that lets the system improve. An architecture with no feedback path is a demo, not a product.
Problem decomposition. Breaking a large task into steps each of which can be independently prompted, evaluated and fixed. Decomposition is what makes a system debuggable.
Business value pillars. Efficiency, transformation, productivity, cost and performance SLAs. Every design decision should trace to one of these or it is unjustified.
Human-in-the-loop placement. Where a person reviews, approves or overrides. Placing it well is a design decision, not an afterthought bolted on after a failure.
Failure-mode budgeting. Deciding in advance what the system does when it is wrong, slow or unavailable — because it will be all three at some point.
Official documentation
Building effective agents ↗ How we built our multi-agent research system ↗ Claude Agent SDK overview ↗ Skilljar: Claude Platform & Solution Design ↗
Exam pro tips
  • When several architectures would work, the exam rewards the simplest one that meets the stated requirements. Multi-agent is rarely the right answer to a question that never mentions parallelism or independent sub-tasks.
  • Watch for scenarios that quietly hand you a constraint — a latency SLA, a compliance boundary, a fixed budget. That constraint is usually the whole answer.
  • Professional-level questions ask you to justify, not just to pick. Practise saying why the rejected option is wrong, not only why yours is right.
  • An architecture with no feedback loop is incomplete. If an option includes evaluation and iteration and the others do not, that is a strong signal.
02

Claude Models, Prompting & Context Engineering

13%

Choosing the right model for the job and controlling its behaviour through prompts and context. Foundations tests the techniques; Professional tests the trade-offs.

Key concepts to master
Model trade-offs. Capability against latency against cost. The most capable model is the wrong answer whenever a smaller one clears the quality bar.
System prompt design. Role, constraints, output contract and refusal behaviour set once, at the top, rather than repeated per request.
Zero-shot, few-shot, chain-of-thought. Escalating techniques. Few-shot buys format reliability; chain-of-thought buys reasoning at the cost of tokens and latency.
Context window management. Deciding what earns a place in a finite window. Everything you include displaces something else.
Prompt caching. Reusing a stable prefix across calls to cut cost and latency. It changes how you order a prompt: stable content first, variable content last.
Modular prompts and Skills. Packaging reusable instructions so behaviour is consistent across a fleet of calls rather than re-invented per feature.
Guardrails in the prompt. Instructions that constrain scope and output shape. Necessary but not sufficient — prompt-based control is advisory, not enforcement.
Progressive disclosure. Supplying context in stages as the task demands it, rather than front-loading everything on the chance it is needed.
Official documentation
Choosing the right model ↗ Prompt engineering overview ↗ Context windows ↗ Prompt caching ↗
Exam pro tips
  • This is the smallest technical domain at 13%, and it overlaps heavily with the Foundations exam. If you hold Architect - Foundations, this is the cheapest domain to refresh rather than relearn.
  • Prompt-based guardrails are advisory. When a question involves a hard requirement, the answer is usually programmatic enforcement, not a firmer instruction.
  • Cost questions almost always have a caching or batching answer hiding in them.
  • Beware options that solve a context problem by buying a bigger window. The exam prefers deciding what does not belong in the window.
03

Integration

19%

The heaviest domain. Connecting Claude to real systems, real data and real auth — and defending the trade-offs when accuracy, latency and security pull against each other.

Key concepts to master
Capability bloat. Handing an agent more tools than the task needs. Every extra tool widens the blast radius and degrades tool-selection accuracy.
Authentication vs. authorisation. Who the caller is versus what they may do. Agents routinely inherit far more authority than the task requires.
RAG pipeline design. Chunking, embedding, indexing and retrieval. Chunking strategy is usually where a weak RAG system actually fails.
Retrieval strategy matching. Matching retrieval to data shape and query pattern — semantic search, keyword, hybrid or structured query. There is no default that always wins.
Connection protocol selection. MCP, direct API/CLI, or agent-to-agent. MCP standardises tool exposure; a direct call is simpler when there is nothing to standardise.
Progressive discovery vs. monolithic context. Letting an agent fetch what it needs as it needs it, rather than pre-loading a large context up front.
Observability at scale. Tracing a request through prompts, tool calls and retries. Without it you cannot diagnose a failure you cannot reproduce.
Accuracy-latency trade-off. Extra verification steps and round trips buy correctness and cost time. The right point on that curve is a business decision, not a technical one.
Official documentation
MCP architecture overview ↗ Implement tool use ↗ Streaming Messages ↗ Skilljar: Enterprise Integration & Production ↗
Exam pro tips
  • At 19% this is the largest domain on the exam. If you are triaging study time, start here.
  • When a scenario mentions an agent with many tools, suspect capability bloat before anything else.
  • Security-flavoured options that narrow scope, shorten credential lifetime or reduce authority are usually correct. Options that add a checking step to an over-privileged design usually are not.
  • RAG questions are more often about chunking and retrieval strategy than about the embedding model.
  • Progressive discovery beats monolithic context whenever the scenario mentions large or unpredictable data volumes.
04

Evaluation, Testing & Optimization

16%

Proving the system works, finding out why it does not, and making it cheaper and faster without quietly making it worse.

Key concepts to master
Evaluation metrics. Accuracy, latency, cost, safety and security. A system optimised for one while blind to the others is not production-ready.
Evaluation datasets. Curated cases including the awkward ones. A dataset made only of easy cases proves nothing.
Mixed evaluation methodology. Combining exact-match checks, model-graded scoring and human review. Each catches what the others miss.
A/B testing. Changing one variable and measuring the difference. Without it, prompt iteration is superstition.
Failure diagnosis. Separating a prompt failure from a retrieval failure from a model mismatch. The fix differs completely for each.
Token and cost optimisation. Caching, batching, model right-sizing and trimming context — applied after you can measure, not before.
Regression risk. An optimisation that improves the average while breaking a specific important case. Aggregate metrics hide this.
Observability-driven iteration. Using production logs and traces to choose what to fix next, rather than intuition.
Official documentation
Define success criteria ↗ Develop tests ↗ Batch processing ↗ Pricing ↗
Exam pro tips
  • The exam consistently rewards measuring before optimising. An option that tunes something without establishing a baseline is usually wrong.
  • Watch for a strong headline accuracy figure masking a weak subgroup — that gap is often the point of the question.
  • Batching trades latency for cost. It is right for asynchronous work and wrong for anything interactive.
  • When a fix is proposed for a failure whose cause has not been identified, that is the trap.
05

Governance, Safety & Risk Management

14%

Designing the safety stack, and placing controls where they actually hold rather than where they are easiest to add.

Key concepts to master
Guardrails and safety controls. Layered constraints on what the system may do. Programmatic enforcement holds where prompt instructions merely suggest.
LLM failure modes. Hallucination, prompt injection, jailbreaks, context loss and silent degradation. You cannot mitigate what you have not enumerated.
Prompt injection. Untrusted content that the model treats as instruction. The mitigation is isolating and fencing untrusted input, not asking the model to ignore it.
Human-in-the-loop validation. A person in the path for high-consequence actions, positioned where a wrong decision is still reversible.
Regulatory compliance. GDPR, HIPAA, FedRAMP and similar regimes constrain where data may live and who may see it, which is an architectural constraint rather than a policy footnote.
Data residency and minimisation. Sending only what the task requires, only where it is permitted to go.
Bias, fairness and transparency. Whether the system treats groups differently, and whether anyone can tell how a decision was reached.
Auditability. Being able to reconstruct after the fact what the system did and on what basis.
Official documentation
Mitigate jailbreaks ↗ Reduce hallucinations ↗ Usage policy ↗ Skilljar: Responsible AI, Safety & Risk for Architects ↗
Exam pro tips
  • Programmatic enforcement beats prompt-based instruction whenever the requirement is hard. This distinction recurs across the whole exam.
  • Prompt-injection answers that rely on telling the model to ignore malicious instructions are wrong. Isolation and least privilege are right.
  • Compliance constraints are architectural. If a scenario names a regime, expect the correct answer to change where data flows, not just what gets logged.
  • Human-in-the-loop is not free. The exam rewards placing it at high-consequence, low-reversibility points rather than everywhere.
06

Stakeholder Communication & Lifecycle Management

14%

The non-engineering domain that decides whether a system ships, gets adopted and survives handover. Weighted the same as governance, and the one architects most often under-prepare.

Key concepts to master
Structured discovery. Interviewing for the real problem rather than the requested solution. The stated ask is frequently not the actual need.
Requirements gathering. Separating must-have from nice-to-have, and surfacing the constraints nobody volunteered.
Architecture decision records. Writing down what was decided, what was rejected and why. The rejected options are the valuable part six months later.
Trade-off communication. Explaining a technical choice in terms of business consequence, to an audience that does not want the implementation detail.
SLA and expectation alignment. Agreeing what good looks like before delivery, in numbers both sides accept.
Handover and documentation. Transferring a system to the people who will run it, with enough context that they can change it safely.
Lifecycle phases. Discovery, design, build, handoff, monitoring and iteration. Architecture work continues well past launch.
Feedback loops with stakeholders. Structured checkpoints that surface drift early, while correcting it is still cheap.
Official documentation
Skilljar: Stakeholder Engagement, Lifecycle & GTM ↗ Customer stories ↗ Transparency ↗ Define success criteria ↗
Exam pro tips
  • This domain and Developer Productivity are the two with no analogue in Architect - Foundations. If you are upgrading from Foundations, this is where your gap is, and it is 14% of the exam.
  • Engineers habitually pick the technically complete answer. Here the correct answer is often the one that manages expectations or surfaces a decision to a stakeholder.
  • Scenarios where a project is technically sound but failing are almost always about communication, expectations or handover.
  • Documentation and decision records are treated as deliverables, not overhead. Options that skip them to move faster are traps.
  • Do not skip this because it is not engineering. It is worth the same as governance and twice as much as developer productivity.
07

Developer Productivity & Operational Enablement

7%

The smallest domain: setting a team up to build with Claude and to operate the system once you have handed it over.

Key concepts to master
Team tooling configuration. Standardising Claude Code settings, permissions and shared configuration so a team behaves consistently rather than per-developer.
Shared project context. Committed configuration such as CLAUDE.md that gives every developer and agent the same working assumptions.
Permission and approval modes. Deciding how much autonomy tooling has in which environment. Sensible defaults matter more than individual discipline.
AI-assisted workflow improvement. Applying Claude to review, test generation and repetitive engineering work, where the payback is measurable.
Operational debugging. Diagnosing a live Claude system from its logs and traces rather than by reproducing it locally.
Enablement and runbooks. Written procedures that let the operating team handle common failures without escalating to the architect.
Official documentation
Claude Code overview ↗ Settings reference ↗ Memory (CLAUDE.md) ↗ Skilljar: Team Enablement & Operational Productivity ↗
Exam pro tips
  • At 7% this is the lightest domain — roughly four or five items. Do not spend study time here at the expense of Integration or Solution Design.
  • Answers favour team-wide, committed configuration over per-developer setup.
  • Enablement questions reward making the operating team self-sufficient rather than keeping the architect in the loop.
  • It is small but not zero. A quick pass over Claude Code settings, permission modes and CLAUDE.md is cheap insurance.
04

Lab & study setup

Professional is a design exam, so the useful practice is designing and defending systems rather than following tutorials. Build one realistic solution end to end and take it further than a demo: evaluate it, threat-model it, document it, and prepare to explain it to somebody who is not an engineer.

⚠ Practise on your own accounts and non-production data

Everything below should be done on personal accounts with material you are free to use. Never point exam practice at production systems, customer data, or corporate networks.

  • Use your own API key and your own billing, not an employer's
  • Practise on synthetic or public documents, never real customer or personal data
  • Keep exam practice off corporate networks and out of production environments
  • Treat anything under NDA as out of bounds for practice material
Subscriptions & access
Personal Anthropic API key (Console, pay-as-you-go)The one thing you genuinely need. Professional is a design exam, but the reasoning about cost, latency, caching and batching only sticks once you have watched real token spend. Set a spending limit before you start.
Personal Claude Max 5x ($100/mo)Worth it if you intend to work through the full prep path and build a solution end to end with Claude Code. Comfortable for a few focused sessions a day without metering anxiety.
Claude Pro ($20/mo)Enough for reading, planning and drafting architecture decision records. Adequate if your hands-on work is mostly design and documentation rather than heavy building.
No subscription at allDefensible for this exam. Nothing on the blueprint requires a paid plan, and the entire official prep path on the Partner Academy is free. Skip the subscription if you already have production Claude experience to reason from.
Your environment
  • A personal Claude account and API key with a spending limit set
  • Claude Code installed, with a committed CLAUDE.md and a deliberate permission mode
  • A scratch repository for architecture notes, decision records and evaluation results
  • A small document corpus you own, for retrieval experiments
  • A note-keeping habit: record why you rejected an option, not only what you chose
Hands-on practice
  • Take one ambiguous business problem and design the full solution: input, processing, output and feedback loop
  • Write an architecture decision record for it, including the options you rejected and the reason
  • Build a small evaluation harness with a dataset that includes awkward and adversarial cases
  • Threat-model the design: enumerate failure modes, then place controls where they actually hold
  • Design the RAG pipeline properly - chunking, indexing and a retrieval strategy matched to the data shape
  • Prepare a stakeholder-facing design review: the trade-offs in business terms, with an SLA you could defend
  • Write the handover runbook for the three most likely production failures
From people who passed
  • Rehearse justifying the simplest adequate architecture. The exam rewards restraint far more than ambition.
  • For every design decision, practise naming the business value it serves. Professional questions ask why, not just what.
  • Do not treat the two non-engineering domains as filler. Together they are 21% of the exam and they are where most architects lose marks.
  • Read every scenario for its constraint - a latency SLA, a compliance regime, a fixed budget. The constraint usually selects the answer.
  • Work the official exam guide's domain list as a checklist and be honest about which task statements you could not currently explain to a colleague.

Sources: Claude Certified Architect - Professional exam guide (official) · Anthropic Partner Academy - certification FAQ · Anthropic / Claude official pricing (consumer plans) · Claude Code setup

05

Official resources

There is no single official “CCAR-P study guide” PDF. The real prep is the Anthropic Academy courses plus the official product docs — everything below is official.

  • Book the exam ↗

    The certification page where you schedule and sit the proctored exam. Sign in first — it's partner-gated, so the page only appears once you're signed in. — CCAR-P certification page

  • Partner portal — exam access ↗

    Where you actually book the exam. Sign in with your partner-organisation account to request the Claude Certified Architect – Professional exam. — claude.com/partners

  • Anthropic Academy — prep path ↗

    The free, self-paced course path this guide's prep section is built from — sign in and enroll. — anthropic-partners.skilljar.com

  • Claude documentation ↗

    The source of truth for exam questions on architecture, integration, evaluation, and governance. — docs.claude.com · code.claude.com/docs

  • All Claude certifications ↗

    The index of all four Claude certifications, with links to every certification page and exam guide. — Partner Academy certifications page

⚠ A word of caution

Skip third-party “free CCAR-P practice test” and brain-dump sites that crowd search results. They are not affiliated with Anthropic, are frequently inaccurate, and dump sites can violate exam terms. Stick to Anthropic Academy and the official docs above.

06

Frequently asked questions

The essentials on eligibility, cost, format, and access — answered from verified facts.

How is this different from Claude Certified Architect - Foundations?

Different exam, not a harder version of the same one. Professional has 63 items to Foundations' 60, costs $175 against $125, and spreads across seven broad domains rather than five deep technical ones. Most importantly it tests Stakeholder Communication & Lifecycle Management (14%) and Developer Productivity & Operational Enablement (7%), which Foundations does not test at all. Foundations goes deeper on Claude Code, MCP and the Agent SDK; Professional goes wider across the whole delivery lifecycle.

Do I need to pass Foundations first?

No. No Claude certification has prerequisites. The exam guide states the credential is awarded on exam performance alone. Foundations is useful preparation, not a gate.

Who is this exam for?

Practitioners in an architect role. The guide recommends 3+ years in systems architecture or platform engineering, at least 6 months hands-on with Claude or comparable LLM systems in production, and experience delivering end-to-end systems from discovery through operationalisation. That experience is recommended, not required.

What is the exam format?

63 multiple-choice and multiple-response items in 120 minutes, each item stating how many responses to select. It is scored on a scaled score of 100-1,000 with 720 to pass, and the score report shows percent-correct by domain.

How much does it cost and how long is it valid?

$175 USD. The credential is valid for 12 months from the date it is awarded. On-time renewal is a free, non-proctored assessment on the Anthropic Partner Academy; a lapsed credential requires retaking the full proctored exam at full fee.

Where do I sit it?

It is proctored - online or at a test centre, per programme policy. Registration and scheduling run through the Anthropic Partner Academy and Pearson VUE.

Can anyone register?

The certification programme runs through the Claude Partner Network, so exam access depends on your organisation's partner status. The prep courses on the Partner Academy are free.

Which domain should I study first?

Integration, at 19%, is the heaviest. After that, Solution Design & Architecture at 17% and Evaluation, Testing & Optimization at 16%. If you already hold Foundations, spend your time instead on Stakeholder Communication and Developer Productivity - together 21% of the exam and entirely absent from Foundations.

How long is the official prep path?

Five courses totalling roughly 12.2 hours: Claude Platform & Solution Design (238 min), Enterprise Integration & Production (158 min), Responsible AI, Safety & Risk for Architects (114 min), Stakeholder Engagement, Lifecycle & GTM (178 min) and Team Enablement & Operational Productivity (45 min).

Is it a coding exam?

No. It is a design and judgment exam. You are not asked to write code, but you are expected to reason about API patterns, integration mechanisms, retrieval design and failure modes with an engineer's precision.

What happens if the exam guide changes?

Anthropic states the guides are subject to change without notice. This page records the version it was checked against, and links the official guide so you can verify any claim at source before you book.