Skip to content

CCAR-P - 6-Week Practice Plan

This plan targets the Claude Certified Architect - Professional exam (CCAR-P). It assumes you are solid on CCAR-F (Foundations) content; Anthropic recommends passing CCAR-F first. If you are not there yet, spend 2 weeks on the Foundations guide first.

Before week 1: register for the exam via the Anthropic Partner Academy and enroll in the free official prep courses there. Booking a date makes the plan real.

Daily commitment: 1.5-2 hours on weekdays, 3-4 hours on one weekend day. Total: roughly 70 hours.

Week 1 - Solution Design and Architecture (Domain 2)

Goal: internalize orchestrator-worker, planner-executor, and single-agent tradeoffs.

  • Read notes/01-advanced-claude-architectures.md
  • Read Anthropic's multi-agent research blog post end to end
  • Reproduce a minimal orchestrator-worker agent: one Opus planner, three Haiku workers, combine into a research summary
  • Instrument the agent to log tokens-per-worker and total cost per trace
  • Convert the same task to a single-agent Sonnet tool loop. Compare cost, latency, quality.
  • Write a one-page decision memo: when would you pick each pattern?
  • Read notes/02-claude-agent-sdk-deep-dive.md
  • Build one minimal example with the Claude Agent SDK (Python or TypeScript)

Deliverable: a repo with both architectures, a cost table, and a decision memo.

Week 2 - Models, Prompting, and Context Engineering (Domain 6)

Goal: master thinking budgets, the memory tool, and context shaping.

  • Read notes/03-extended-thinking-and-context-management.md
  • Read the extended thinking docs cover to cover
  • Take a task where Sonnet scores 70% without thinking. Turn on thinking with budgets of 1K, 4K, 16K. Measure quality and cost.
  • Implement interleaved thinking in a 5-tool agent. Confirm thinking blocks appear between tool calls.
  • Build a simple client-side memory tool integration. Persist user preferences across sessions.
  • Take a 180K-token RAG prompt and cut it to 60K tokens without quality loss via summarization, caching, and retrieval.
  • Compute: what is the break-even cache read count for a 30K-token cached prefix?

Deliverable: a markdown write-up of thinking budget ROI curves and a before/after context diet.

Week 3 - Integration: Tool Use and MCP (Domain 1)

Goal: design production-grade tools and MCP servers. Integration is the heaviest domain at 19%.

  • Read notes/04-tool-use-and-mcp-integration.md
  • Audit a 25-tool agent. Prune, merge, or dispatcher-ify until under 20.
  • Add explicit examples to three tool descriptions and A/B against the old descriptions. Measure tool-selection accuracy.
  • Build an MCP server over streamable HTTP. Add OAuth 2.1 auth. Deploy behind a reverse proxy.
  • Add the code execution tool to an agent. Run a data analysis task end to end.
  • Experiment with parallel tool use. Confirm Claude emits multiple tool_use blocks in one response.
  • Read the MCP specification sections on transports and authentication

Deliverable: a deployable MCP server with auth, versioning, and a README.

Week 4 - Evaluation, Testing and Optimization (Domain 3)

Goal: ship evals and tracing that catch regressions before users do, then squeeze cost.

  • Read notes/05-evaluation-and-observability.md
  • Read notes/06-cost-latency-optimization-at-scale.md
  • Build a 50-item eval dataset for one real task
  • Implement an LLM-as-judge grader with a structured rubric. Calibrate against 10 human-labeled items.
  • Wire a tracing system (OpenTelemetry, Langfuse, or custom) that logs tokens, tool calls, cache reads, thinking tokens per trace
  • Add a regression gate in CI: block merges that drop quality by more than 3% on the eval set
  • Add prompt caching breakpoints to your agent. Measure cache hit rate over 100 requests.
  • Move one workload to the Batch API. Verify the 50% discount and 24h SLA fit.
  • Implement model routing: Haiku triage, Sonnet execute, Opus fallback on low confidence.

Deliverable: an eval harness, a regression gate, and a cost-optimization write-up.

Week 5 - Enterprise Deployment and Governance (Domains 2 and 4)

Goal: deploy on Bedrock and Vertex, then wrap the deployment in governance controls.

  • Read notes/07-enterprise-deployment-bedrock-vertex.md
  • Read notes/08-governance-safety-and-risk-management.md
  • Deploy the same agent via Amazon Bedrock. Handle IAM auth, regional model IDs.
  • Deploy the same agent via Google Vertex AI. Handle service account auth, publisher model IDs.
  • Enable PrivateLink (Bedrock) or VPC Service Controls (Vertex) and confirm no public egress.
  • Document a zero-data-retention (ZDR) request flow if your workload requires it.
  • Run a small red-team suite (jailbreaks, prompt injections, PII leaks). Document failures and mitigations.
  • Add an input classifier (Haiku) in front of an Opus agent. Measure the block rate and latency overhead.
  • Add a human-approval gate (pre-tool hook) on one irreversible action. Log the approval in the audit trail.
  • Write a one-page model risk entry for your agent: owner, purpose, risk tier, validation evidence, review cadence.
  • Map your agent against SOC 2, GDPR, and HIPAA: which controls are yours, which are the provider's?

Deliverable: a deployment matrix, a red-team report, and a governance memo.

Week 6 - Stakeholder Communication, Lifecycle, and Exam Readiness (Domain 5 + review)

Goal: cover the business-facing domain, then close gaps and build exam stamina.

  • Read notes/09-stakeholder-communication-and-lifecycle-management.md
  • Write a one-page executive brief for your week-1 agent: value, cost per outcome, risks, rollout phases
  • Define go/no-go gates for a pilot-to-production promotion: quality, safety, cost, operations thresholds
  • Write a migration plan for a model version deprecation: eval diff, phased rollout, rollback
  • Draft a runbook covering three alerts: quality drift, cost spike, provider outage
  • Re-read all notes in order, annotating anything unclear
  • Work through every scenario in scenarios.md timed (about 2 minutes per question)
  • Do the practice question bank untimed, then re-do misses timed
  • Re-read fact-sheet.md and flag any fact you cannot explain in your own words
  • Re-read strategy.md the night before
  • Confirm current model IDs, pricing, and your Pearson VUE appointment logistics

Deliverable: ready to sit the exam.

Hands-On Exercises (Pick Any Two)

  1. Build a research agent: orchestrator-worker, 20 parallel Haiku workers, consolidated summary via Opus. Target: under $0.10 per research run.
  2. Build a customer support agent with memory tool, extended thinking on escalations, and an eval harness tracking CSAT predictions.
  3. Build a code review agent using the Agent SDK with hooks that block commits failing quality checks.
  4. Build a compliance redaction pipeline via Batch API that processes 10K documents nightly at under $50.
  5. Build a multi-cloud Claude gateway that routes between first-party, Bedrock, and Vertex based on request attributes.
  6. Write the full launch package for any of the above: executive brief, go/no-go gates, runbook, and enablement plan (Domains 4 and 5).

Red Flags - Do Not Sit the Exam Yet If

  • You cannot explain when orchestrator-worker beats a single agent in cost
  • You cannot wire extended thinking with preserved signatures across tool calls
  • You have never deployed Claude via Bedrock or Vertex
  • You do not have an eval harness for at least one production-like task
  • Your cache hit rate on your own projects is under 50% for repeatable workloads
  • You cannot draw the MCP client-server-transport diagram from memory
  • You cannot explain ZDR, BAA coverage, or where prompt injection defense actually lives (the tool boundary)
  • You cannot write a go/no-go gate or a model deprecation migration plan from memory