CCAR-P exam topics
The CCAR-P exam (Claude Certified Architect (Professional)) covers 7 domains and 38 topics. This page lists all of them with the weight of each domain, so you can see what the exam tests and where most of the marks are.
Domains and weights
Domain 1: Solution design & architecture · 17%
- Translating Business Problems into Claude Solutions
Translating a business problem into a Claude solution produces a measurable success criterion, a bounded scope, and a defensible boundary between the model and ordinary software.
- End-to-End Architectures and Feedback Loops
Every production Claude system has six stages: Input — ingestion, validation, normalisation, PII handling (decide before model); Context assembly — retrieval, memory, prompt construction; Processing — model call(s): single, workflow, or agent loop; Output — schema validation, business-rule checks, guardrail screening, designed failure path. Model output is untrusted until validated.; Delivery — sync, async, event, or draft-for-review (risk decision); Feedback loops — user feedback, automated eval on sampled traffic, monitoring → iteration.
- Selecting Architectural Patterns
Find the simplest solution that meets requirements; add complexity only when it demonstrably improves outcomes. Escalate tier only when the simpler tier cannot meet a named requirement.
- Multi-Agent Systems and Orchestration
Excels at breadth-first work that parallelises cleanly: research across independent directions, analysis of many independent units, tasks exceeding one context window, work needing distinct specialisms/tools/permissions.
- Decomposition Techniques for Complex Problems
One pass covering too much → inconsistent depth, missed items, contradictory judgements. Fix (both mandatory):; Per-unit local passes (consistent depth on every unit); Separate integration pass (cross-unit issues invisible to local passes).
- Aligning Solutions to Business Value Pillars
Test: does the capability already exist? Yes → efficiency. No → transformation. Design Decisions Tied to Pillars: Model selection — per-task cost/performance; smallest model that meets quality bar; Prompt caching — ~90% cache-read cost reduction; serves cost + latency SLA (static first, dynamic last); Batch processing — halves cost of async work; sacrifices request-path latency; Multi-agent — ~15× tokens; name the pillar that pays for it.
Domain 2: Claude models, prompting & context engineering · 13%
- Selecting Claude Models on Trade-Offs
Model selection = which model clears this task’s quality bar at the lowest cost that meets the latency budget. Never “which model is best”.
- System Prompts, Templates, and Guardrails
System prompt = standing contract (role, audience, policy, output format). User turns carry per-request data. Template = prompt with named slots. Guardrail = limit on what the model receives or returns.
- Zero-Shot, Few-Shot, and Chain-of-Thought
Diagnose before adding anything. Instruction quality is the first, free fix. Matching Technique to Failure: Output format wrong/drifting, Inconsistent classification at boundaries, Wrong answers on multi-step reasoning, Stated rules ignored, Any of the above.
- Optimising Context Windows and Token Usage
Context window = everything Claude sees in one request (system, history + tool results, tool definitions, output + thinking). Every token is billed every time it appears.
- Prompt Reuse: Caching, Modular Prompts, and Skills
Three mechanisms for three reuse questions: Prompt caching, Modular prompts, Skills; Prefix Mechanics: Order: tools → system → messages. Front-load stability: static system/policy/examples first, dynamic content last. A timestamp early in the prompt defeats the entire cache.
Domain 3: Integration · 19%
- Capability Bloat and Least-Privilege Configuration
Least privilege = agent gets only the tools its role needs. Capability bloat = every tool beyond that. Removal beats mitigation every time.
- Authentication, Authorisation, and the Gaps Between
Authentication = who is calling. Authorisation = what may they do. A security gap exists wherever effective privilege exceeds the requesting user’s legitimate need.
- Accuracy–Latency Trade-offs
Every integration has a latency requirement (e.g. p95) and a quality requirement. The skill is holding both at once.
- Observability and Monitoring at Scale
LLM systems break anecdote debugging: non-deterministic, so a single re-run proves nothing. A green uptime dashboard says nothing about answer quality.
- RAG Pipeline Design: Chunking and Indexing
RAG stages: ingest → chunk → embed → index → retrieve → (rerank) → generate. Generation quality is capped by retrieval quality.
- Matching Retrieval to Data Shape and Query Pattern
Retrieval strategy rests on two axes: data shape and query pattern. Then name the failure that matters most and pick the cheapest strategy whose failure mode is acceptable.
- Choosing the Integration Mechanism
Four mechanisms connect Claude to surrounding systems: MCP server, Direct API/SDK, Headless (SDK/CLI), Agent-to-agent.
- Progressive Discovery vs Monolithic Context
Two ways to load tools, procedures and knowledge: Monolithic — load everything upfront (simplest; best for small stable surfaces; cacheable); Progressive discovery — load each piece on demand (Skills, deferred tool loading, agentic retrieval).
Domain 4: Evaluation, testing & optimisation · 16%
- Defining Evaluation Metrics
Every metric has three parts: definition, measurement method, and target with baseline. Define metrics before the build.
- Designing Evaluation Datasets and Test Frameworks
Non-deterministic outputs → report SEM and CIs; use clustered SEs for related questions; resample and average; run power analysis before the eval. High-stakes/high-variance behaviours need more cases.
- A/B Testing and Iterative Improvement
Only an isolated change can be credited. Land the first, establish the new baseline, then test the second. “Since we’re touching it anyway” turns experiments into deployments.
- Diagnosing Prompt Failure, Hallucination, and Model Mismatch
Diagnosis finds why a system fails before changing anything. Order: Trace and reproduce (pattern across cases); Isolate the stage (evidence first); Classify against signatures; Run the cheapest decisive test; Fix at the failing layer; Mint regression cases.
- Optimising Token Usage, Latency, and Cost-Performance Trade-offs
Key Concept: Optimisations move along the frontier — they do not repeal it. Name the spend and the saving, with eval evidence the spend is affordable.
- Monitoring System Performance in Production
Launch-time evals prove readiness once; monitoring proves it stays true. Domain 3 built the plumbing; Domain 4 runs the quality practice on top.
Domain 5: Governance, safety & risk management · 14%
- Guardrails, Safety Controls, and the Surface a Risk Belongs To
A guardrail is a control that stops a Claude system causing harm. Each risk belongs on the one surface that can hold it.
- Risks, Limitations, and the Failure Modes That Stay Silent
A risk register is an architecture artefact. Every row has five fields: Mechanism — what goes wrong; Blast radius — how far, who bears it, whether undoable; Detection signal — what moves in the telemetry; Owning control — which 5.1 surface carries it; Residual, and who accepted it.
- Human-in-the-Loop Validation: Placement, Pattern, and Proof
Placement comes before pattern. The gate belongs where consequence is high and the action is hard to reverse.
- Regulatory Compliance as an Architecture Constraint
An obligation becomes an architecture constraint when it changes what you build. Scope is the whole data path, not the application.
- Ethical AI: Bias, Fairness, and Transparency
Bias is a gap in outcomes between groups. Dropping the protected attribute proves nothing — proxies carry the signal.
Domain 6: Stakeholder communication & lifecycle management · 14%
- Structured Discovery and Requirement Gathering
The request is not the requirement. Discovery finds the problem behind the request. Six Outputs Discovery Must Produce: Business problem with measured baseline — hours, error rates, backlog, revenue at risk; Real workflow (including workarounds); Output the system must produce (fixes task shape); Consequence when that output is wrong + thresholded, owned success criteria; Data — what exists, quality, ownership, lawful basis; Constraints and accountable decision-maker.
- Communicating Architectural Decisions and Trade-offs
Simplifying past the point of truth is not translation. Say the Error Rate Out Loud: State the rate, the population, and the handling path for residual errors. A demo is a sample from the favourable tail — never a substitute for the measured rate.
- Expectation Alignment and Service-Level Commitments
Engineer first, then promise. Never the reverse. Anatomy of a Defensible Commitment (Six Elements): Metric — what is measured and how; Percentile or threshold — means conceal the tail; Measurement point — where the clock starts and stops; Window — over what period compliance is assessed; Population — who and what it covers (defined before measurement); Breach behaviour — what happens when it is missed.
- Documentation and Implementation Guidance
A handover pack transfers judgement, not history. Test for every page: does it change what a competent stranger would do?
- Supporting the Lifecycle Phases
Phases are gated by outputs, not dates. The architect’s job does not end at deployment. Phase Outputs: Discovery, Design, Build & validate, Handoff, Operate & monitor, Iterate.
Domain 7: Developer productivity & operational enablement · 7%
- Team Tooling Configuration: Guidance Versus Enforcement
One question sorts everything: does an element shape output the team collectively owns? If yes → shared and version-controlled. If no → personal.
- AI-Assisted Developer Workflows: Measure the Work, Package the Practice
Outcome measures (properties of the work): cycle time, defect escape rate, review latency, time for a new joiner to make a safe change, incident time-to-resolution.
- Debugging and Operational Resolution: Evidence Before Change
When a production agent behaves unexpectedly, the first move is to read what happened: inputs, retrieved context, tool calls and results, outputs, versions in play.
CCAR-P exam facts
| Questions | 63 |
|---|---|
| Time limit | 120 minutes |
| Passing score | 720 on a scale of 1,000 |
| Exam fee | $175 USD |
| Certification valid for | 12 months |
Study these topics
- Read the free CCAR-P study guide: one lesson per topic above, each with its sources in Anthropic's documentation.
- See how to prepare for CCAR-P for a study plan ordered by domain weight.
- Test yourself with the free 10-question CCAR-P mock test, or the full CCAR-P practice exam.
Skillmarked is independent and not affiliated with Anthropic. The topic list follows the public exam outline; check the official exam page for the current version before you book.