Skip to content

CCDV-F - 5-Week Practice Plan

Daily commitment: 1.5-2 hours weekdays, 3-4 hours one weekend day. Total: ~55 hours.

This plan is hands-on. Every week ends with code you can run. Weeks are ordered by the exam blueprint: Applications and Integration (33.1%) gets the most time, Claude Code and evals (under 6% combined) get a focused day, not a week.

Week 1 - Messages API and Streaming (Domain 1)

  • Read notes/01-claude-api-fundamentals.md and notes/02-messages-api-and-streaming.md
  • Set up ANTHROPIC_API_KEY in your environment
  • Install the Python SDK and TypeScript SDK
  • Send a basic Messages request from both SDKs
  • Inspect the response: content blocks, usage, stop_reason
  • Implement streaming. Print tokens as they arrive.
  • Catch and reassemble all SSE event types: message_start, content_block_start/delta/stop, message_delta, message_stop
  • Handle a pause_turn scenario by logging it
  • Write a 30-line CLI that takes a prompt and streams the response

Deliverable: a CLI that streams from the current Sonnet model.

Week 2 - Tool Use and MCP (Domains 5 and 3)

  • Read notes/03-tool-use-function-calling.md
  • Define a single tool (e.g., get_weather) and call it via auto choice
  • Walk through the tool_use / tool_result lifecycle by hand
  • Implement parallel tool use: define two independent tools and confirm Claude calls both in one turn
  • Use structured output (schema-constrained) for JSON extraction
  • Implement an agent loop that runs until end_turn with an iteration cap
  • Add a structured error response (is_error: true) and observe Claude recovering
  • Try tool_choice: any and tool_choice: {type: "tool", name: "X"}
  • Connect one MCP server (local or remote) and call one of its tools; skim the MCP architecture at modelcontextprotocol.io

Deliverable: a tool-use agent with at least 3 tools, including one error path and one MCP-sourced tool.

Week 3 - Caching, Batch API, Files, Model Selection (Domains 2 and 1)

  • Read notes/04-prompt-caching-and-batch-api.md, notes/05-files-api-citations-and-pdfs.md, and notes/08-model-selection-and-optimization.md
  • Add cache_control to a 30K-token system prompt; verify cache_creation_input_tokens and cache_read_input_tokens in usage
  • Compute cache ROI on a synthetic 100-request workload
  • Try the 1-hour cache TTL and reason about when it beats the 5-minute default
  • Submit a batch of 100 requests via the Batch API; poll until complete; retrieve results by custom_id
  • Handle per-item batch errors gracefully
  • Upload a PDF via the Files API and reference it by file_id in two requests
  • Enable citations on a multi-document request and inspect the citation spans
  • Run the same classification task on Haiku, Sonnet, and Opus; compare cost, latency, and accuracy
  • Build a tiny model router: cheap model for easy inputs, capable model for hard ones
  • Use the token counting endpoint to pre-price a large prompt

Deliverable: a "doc Q&A" app with cached system prompt, PDF input, citations, and a cost comparison table across model tiers.

Week 4 - Agents, Error Handling, Security (Domains 3, 6, and 1)

  • Read notes/09-agents-and-workflows.md and notes/06-error-handling-rate-limits-retries.md
  • Implement two workflow patterns by hand: prompt chaining and routing
  • Extend your Week 2 agent: iteration cap, token budget ceiling, circuit breaker on repeated identical tool calls
  • Add a human-in-the-loop gate on one destructive tool
  • Implement exponential backoff with jitter for 429, 500, 529; honor retry-after
  • Distinguish retryable from non-retryable errors with typed SDK exceptions
  • Read the security half of notes/10-security-safety-claude-code-and-evals.md
  • Red-team your own agent: put an injection instruction inside a tool result and watch what happens; then add defenses
  • Validate one tool's model-supplied path/parameter as untrusted input
  • Add structured logging: model, tokens, latency, stop_reason, tool calls

Deliverable: a hardened agent with retries, budgets, an approval gate, and a documented injection test.

Week 5 - Claude Code, Evals, Review (Domains 7, 8, and review)

  • Read the Claude Code and evals parts of notes/10-security-safety-claude-code-and-evals.md and notes/07-sdks-python-typescript-and-cli.md
  • Use Claude Code on a real repo: write a CLAUDE.md, adjust permissions in settings, add one hook, configure one MCP server
  • Run Claude Code headless (claude -p) once
  • Build a 20-case eval set for your Week 3 app: exact-match grader for the classification path, LLM-as-judge with a rubric for one open-ended path
  • Run the eval suite before and after a prompt change; record the score delta
  • Compare Python sync, Python async, and TypeScript SDK ergonomics for the same feature
  • Re-read all notes and the fact sheet
  • Walk through every scenario in scenarios.md under timed conditions
  • Work the practice questions
  • Re-read strategy.md the night before
  • Verify current model IDs, pricing, and your Pearson VUE appointment logistics

Deliverable: ready to sit the exam.

Hands-On Project Ideas (Pick One)

  1. PDF Q&A bot: upload via Files API, cache system prompt, citations enabled, streaming UI.
  2. Structured extractor: read invoices/contracts and produce schema-validated JSON.
  3. Backfill enrichment job: 10K records via Batch API with cache, idempotent retries.
  4. Customer support agent: 5-tool agent with routing, parallel tool use, approval gate, hardened retries.
  5. Eval harness: version-controlled test set with exact-match, code, and LLM-judge graders wired into CI.

Red Flags - Do Not Sit the Exam Yet If

  • You cannot describe every SSE event type and its payload
  • You have never used cache_control and verified the usage accounting
  • You have never run a tool loop end to end
  • You have never used the Batch API
  • You cannot name the five workflow patterns or say when an agent is the wrong choice
  • You cannot explain the cache write/read economics or the model-routing cost lever
  • You cannot describe two prompt-injection defenses beyond "tell the model to be careful"
  • Your retry logic does not honor retry-after
  • You do not know what CLAUDE.md, hooks, and an LLM-as-judge grader are