Claude Certified Architect – Foundations (CCAR-F)

Learning plan

About 45 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 5 weeks

What the exam covers

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

Before you start

• Eligibility: the exam is open only to people at organizations in the Claude Partner Network. You register through the Anthropic Partner Academy, then schedule with Pearson VUE
• A Claude API key with a few dollars of credit, Claude Code installed locally, and working Python or TypeScript
• Anthropic recommends 6+ months of hands-on building with the Claude API, Agent SDK, Claude Code and MCP

1. Weeks 1–2 · 12 h

   Agentic loops and multi-agent orchestration (largest domain)

   • Write an agentic loop that keeps going while stop_reason is "tool_use" and stops on "end_turn", appending tool_result blocks to history each turn
   • Explain hub-and-spoke orchestration. Subagents get isolated context, so the coordinator must pass findings explicitly in the Task prompt
   • Configure AgentDefinition (description, system prompt, tool restrictions), include "Task" in allowedTools, and spawn parallel subagents in one response
   • Choose programmatic enforcement (PreToolUse/PostToolUse hooks, prerequisite gates) over prompt instructions when compliance must be guaranteed
   • Use --resume with named sessions and fork_session, and know when a fresh session with an injected summary beats resuming stale tool results

   Hands-on: With the Agent SDK, build a support agent with get_customer, lookup_order, process_refund and escalate_to_human tools. Add a hook that blocks process_refund until get_customer returns a verified ID and redirects refunds over $500 to escalation. Then send it a message with three separate complaints and check it decomposes them.

   Use: Claude Agent SDK overview · Building effective agents (Anthropic engineering) · Introduction to subagents (Anthropic Academy)

2. Week 2 · 8 h

   Tool design and MCP integration

   • Rewrite near-duplicate tool descriptions (e.g. analyze_content vs analyze_document) with input formats, example queries and boundaries so selection is reliable
   • Return structured MCP errors using isError plus errorCategory (transient, validation, permission, business) and isRetryable, and tell an access failure apart from a valid empty result
   • Set tool_choice to "auto", "any" or a forced {"type":"tool","name":...} and say which one guarantees a tool call
   • Scope MCP servers correctly, with project .mcp.json (env vars like ${GITHUB_TOKEN}) for team servers and ~/.claude.json for personal ones. Know when MCP resources beat exploratory tool calls
   • Pick the right built-in tool (Grep for content, Glob for paths, Edit for unique-anchor changes, Read+Write as the fallback)

   Hands-on: Write a local stdio MCP server exposing two similar search tools and one resource that lists available documents. Register it in a project .mcp.json with an environment-variable token, ask ambiguous questions in Claude Code, and fix the descriptions until tool choice is consistent.

   Use: Introduction to Model Context Protocol (Anthropic Academy) · Model Context Protocol: Advanced Topics (Anthropic Academy) · Tool use with Claude (docs) · Connect Claude Code to tools via MCP

3. Week 3 · 9 h

   Claude Code configuration and CI/CD

   • Lay out the CLAUDE.md hierarchy (user, project, directory), use @import, and check loaded files with /memory
   • Write .claude/rules/ files with paths globs (e.g. "**/*.test.tsx") and explain why they beat subdirectory CLAUDE.md files for scattered test files
   • Build a project skill in .claude/skills/ with context fork, allowed-tools and argument-hint frontmatter, and a shared slash command in .claude/commands/
   • Decide between plan mode and direct execution, and use the Explore subagent to keep verbose discovery out of the main context
   • Run Claude Code headless in CI with -p, --output-format json and --json-schema, and review with a separate instance from the one that wrote the code

   Hands-on: In a personal repo, add a project CLAUDE.md, two path-scoped rules files, a forked skill and a /review command. Then wire a GitHub Actions job that runs claude -p with --output-format json and fails the build when the JSON reports a high-severity finding.

   Use: Claude Code in Action (Anthropic Academy) · Claude Code memory and CLAUDE.md docs · Claude Code skills docs · Claude Code GitHub Actions

4. Weeks 3–4 · 10 h

   Prompting, structured output and context reliability

   • Replace vague instructions ("be conservative") with explicit report/skip criteria, and write 2-4 few-shot examples that show reasoning on ambiguous cases
   • Extract data via tool_use with a JSON schema that uses nullable fields, "unclear" enums and "other" + detail, then run a validation-retry loop that feeds back the specific error
   • Know the Message Batches API facts: 50% cheaper, up to a 24-hour window, custom_id correlation, no multi-turn tool calls. Use it for overnight jobs, never pre-merge checks
   • Keep a persistent case-facts block, trim verbose tool output, and counter lost-in-the-middle by putting key findings first
   • Escalate on explicit human requests, policy gaps and no progress, not on sentiment or self-reported confidence. Calibrate review with stratified sampling

   Hands-on: Build an invoice extractor that defines an extraction tool with required and nullable fields, validates that line items sum to the total, and retries with the validation error. Then submit 50 documents through the Message Batches API and resubmit only the failed custom_ids.

   Use: Prompt engineering overview (Claude docs) · Structured outputs (Claude docs) · Batch processing (Claude docs) · Effective context engineering for AI agents · Claude Cookbooks on GitHub

5. Weeks 4–5 · 6 h

   Exam guide scenarios, sample questions and review

   • Work all 12 sample questions in the official guide and explain in one sentence why each wrong option is wrong
   • For each of the 6 scenarios (support agent, Claude Code codegen, multi-agent research, developer productivity, CI, data extraction), list the task statements it tests
   • Re-read the out-of-scope list (fine-tuning, vision, streaming, computer use, cloud-provider config) so you don't over-study

   Use: CCAR-F official exam guide (PDF) · CCAR-F prep courses (Anthropic Partner Academy)

You're ready to book when…

• You can write a stop_reason-driven agent loop and a coordinator that spawns parallel subagents from memory
• You can say where each config lives (.claude/commands vs ~/.claude/commands, .mcp.json vs ~/.claude.json, .claude/rules paths) without looking it up
• You answer all 12 official sample questions correctly and can justify every distractor
• You can pick hooks vs prompts, batch vs synchronous, and plan mode vs direct execution for any scenario in under 30 seconds

Exam tips

• Each sitting shows 4 scenarios picked at random from 6, so prepare all six. Multiple-response items state how many answers to select.
• The guide's sample answers consistently favour the proportionate fix (better tool descriptions, explicit criteria, programmatic gates) over new infrastructure like classifiers or routing layers.
• Retakes need waits of 14, 30, then 90 days, with at most 4 attempts per exam in a rolling 12 months. On-time renewal is a free, non-proctored assessment.