Claude Certified Architect (Foundations) — CCAR-F
Complete Study Course
Exam Code: CCAR-F
Who This Exam Is For
Solution architects and technical practitioners designing Claude-powered agentic systems.
No formal prerequisite or experience bar is set by Anthropic, but questions assume hands-on building experience with Claude (roughly ~6 months recommended). Credential is awarded purely on exam performance.
Domain Blueprint (Weighted)
| Domain | Weight | Lessons | Focus |
|---|---|---|---|
| 1. Agentic Architecture & Orchestration | 27% | 7 | Agentic loops, orchestration patterns, subagents, guardrails, Agent SDK, task decomposition, session state |
| 2. Tool Design & MCP Integration | 18% | 5 | Tool schemas, structured errors, tool distribution/choice, MCP servers/clients, built-in tools |
| 3. Claude Code Configuration & Workflows | 20% | 6 | CLAUDE.md hierarchy, slash commands/skills, path-specific rules, plan mode, iterative refinement, CI/CD |
| 4. Prompt Engineering & Structured Output | 20% | 6 | System prompts, few-shot, structured output via tools, validation/retry loops, batch processing, multi-pass review |
| 5. Context Management & Reliability | 15% | 6 | Context windows, caching, escalation/ambiguity, error propagation, codebase exploration, human review, provenance |
On Skillmarked: 30 free lessons (one per task statement) and 30 practice questions, with a practice scenario after every lesson.
Recommended Study Plan (15–25 hours with prior Claude experience)
Week 1 – High-weight foundations (Domains 1 + 3 ≈ 47%)
- Master the agentic loop lifecycle (
stop_reason:tool_usevsend_turn) - CLAUDE.md hierarchy, scoping, skills vs commands
- Build a small multi-tool agent loop of your own
Week 2 – Judgement domains (Domains 4 + 2)
- Prompt patterns that appear in production
- Tool schema design and MCP integration
- “Simplest fix that works” mindset
Week 3 – Reliability + mocks (Domain 5)
- Context window strategies, caching, error propagation
- Take diagnostic → target weak domains
- Full-length timed mocks (when ready)
Critical Concepts (High-Yield)
Agentic Loop (Domain 1.1)
The loop is code-controlled, not prompt-controlled:
- Send request (history + system + tools)
- Inspect
stop_reason - If
"tool_use"→ execute tools, appendtool_resultmessages, continue - If
"end_turn"→ finish
Anti-patterns the exam punishes:
- Parsing natural language (“I’m done”)
- Content-type checks (
response.content[0].type == "text") - Arbitrary iteration caps as primary stop condition
CLAUDE.md hierarchy (Domain 3): Project > User > Global; skills live under .claude/skills/ or .claude/commands/.
Context management: Prefer prompt caching for static system content; handle long conversations with summarisation / subagents / progressive disclosure.
Official Registration
Anthropic Partner Academy (currently partner organisations):
https://anthropic-partners.skilljar.com/page/partner-certifications