Claude Certified Architect – Professional (CCAR-P)

Learning plan

About 80 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 8 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• Eligibility: only people at Claude Partner Network organizations can register, through the Anthropic Partner Academy, with delivery by Pearson VUE
• Anthropic recommends 3+ years in systems architecture or platform engineering and 6+ months running Claude (or a comparable LLM) in production. This is recommended, not required
• Comfort with the CCAR-F material (agent loops, MCP, Claude Code configuration). The Professional exam assumes it and moves up to enterprise design trade-offs

1. Weeks 1–2 · 18 h

   Solution architecture, model choice and prompt design

   • Map a business problem to the right pattern (single augmented-LLM call, fixed workflow, or autonomous agent) and justify it on cost, latency and SLA
   • Design multi-agent decomposition and say when an orchestrator-worker design pays for its added cost
   • Choose between Opus, Sonnet and Haiku tiers and extended thinking for a stated quality/latency/cost budget
   • Put static system prompts and policy documents ahead of dynamic content so prompt caching cuts both time-to-first-token and cost
   • Use Skills and modular prompts for reuse, and know zero-shot vs few-shot vs chain-of-thought trade-offs

   Use: Building effective agents (Anthropic engineering) · Claude models overview · Prompt caching (Claude docs) · Claude Platform 101 (Anthropic Academy)

2. Weeks 2–4 · 18 h

   Integration, covering RAG, protocols, auth and observability (largest domain)

   • Design a RAG pipeline (chunk size, overlap, hybrid BM25 + embedding retrieval, re-ranking) matched to the data shape and query pattern
   • Choose MCP, a direct API/CLI call, or agent-to-agent for a given integration, and compare progressive discovery with monolithic context loading
   • Audit an agent's tool list for capability bloat and apply least privilege by removing tools rather than just logging them
   • Find authentication and authorization gaps (per-user scopes, secrets handling, service identities) in a proposed design
   • Choose tracing and monitoring for multi-step agents at scale

   Hands-on: Build a RAG service over a public doc set (for example a product's docs). Compare fixed-size and heading-based chunking, cache the system prompt, and log retrieved chunk IDs alongside every answer so you can trace bad answers back to retrieval.

   Use: Claude Cookbooks (RAG, tool use, evals) · Model Context Protocol documentation · Effective context engineering for AI agents

3. Weeks 4–5 · 14 h

   Evaluation, testing and optimization

   • Define success metrics across accuracy, latency, cost, safety and security before building, as Anthropic's eval guidance recommends
   • Build an eval set that mixes code-graded, model-graded and human-graded checks, and run A/B comparisons of prompt or model changes
   • Diagnose whether a failure is the prompt, retrieval (stale or irrelevant chunks), or a model mismatch
   • Cut token spend with caching, Message Batches and smaller models without dropping below the quality bar

   Hands-on: Write a 50-case eval for your RAG service with expected answers. Score two prompt variants and two model tiers on accuracy, p95 latency and cost per 1,000 queries, then write a one-paragraph recommendation as you would for a client.

   Use: Define success criteria and build evals (Claude docs) · Batch processing (Claude docs)

4. Weeks 6–7 · 18 h

   Governance, stakeholders and operational enablement

   • Layer guardrails (input screening, tool permissions, hooks, human approval) and list common LLM failure modes and mitigations
   • Apply GDPR, HIPAA and FedRAMP constraints to data flow, retention and hosting choices, and address bias, fairness and transparency
   • Run structured discovery, document an architecture decision record, and set SLAs and expectations with non-technical stakeholders
   • Support the lifecycle from discovery and design through handoff, monitoring and iteration
   • Roll out Claude Code to a team (shared CLAUDE.md, settings, MCP servers) and help debug operational issues

   Use: Claude Code overview · Claude Code hooks · CCAR-P prep courses (Anthropic Partner Academy)

5. Weeks 7–8 · 12 h

   Blueprint self-assessment and review

   • Score yourself 1-5 against every objective in the 7-domain blueprint and re-study anything below 4
   • Work the guide's sample items on least privilege, prompt caching and RAG diagnosis, and write your reasoning before checking the key
   • Spend extra review time on Integration (19%) and Solution Design (17%), which together are over a third of the exam

   Use: CCAR-P official exam guide (PDF)

You're ready to book when…

• You can defend a workflow vs agent vs single-call decision for a new use case in terms of cost, latency and risk
• You have shipped one end-to-end Claude solution with RAG, an eval suite and logging, as the guide recommends
• You can name the right control (remove tool, hook, human-in-the-loop, data residency) for a given compliance or safety risk
• You rate yourself 4/5 or higher on every blueprint objective

Exam tips

• The exam has 63 multiple-choice and multiple-response items in 120 minutes, and you need a scaled 720/1000. Your score report shows percent correct by domain.
• New exam content goes live on December 7, 2026, with updated guides published November 17, 2026. Check which version your booking date falls under.
• If you fail, retake waits are 14, 30, then 90 days, with up to 4 attempts in a rolling 12 months.