How to Pass the Claude Certified Architect – Foundations Exam
A domain-by-domain study plan for the Claude Certified Architect – Foundations exam (CCA-F): what the 4-of-6 scenario format really tests, and where to spend your prep time.
The Claude Certified Architect – Foundations exam is not a coding test. It is a design test — 60 questions drawn as four scenarios from a pool of six, each asking you to defend a choice about orchestration, isolation, reliability, or cost. If you prepare like it's a syntax quiz, you'll run out of time on the reasoning.
Here's how to prepare for what it actually measures.
Know the format before you study
Four scenarios, 15 questions each, 120 minutes, scored 100–1000 with a 720 pass. The pool of six is public knowledge, and that changes your strategy completely:
Because the exam draws four scenarios from six, preparing every scenario means you can afford to lose your single weakest one and still clear 720. Skipping scenarios is the most common self-inflicted wound.
Where the weight actually sits
Not all domains are equal. Spend your hours in proportion to the blueprint:
| Weight | What it tests | |
|---|---|---|
| Orchestration | 27% | Agentic architecture, subagents, context isolation |
| Claude Code | 20% | Hooks, commands, Agent SDK, CLAUDE.md scope |
| Prompt & Output | 20% | Few-shot, structured output, enforcement |
| Tool / MCP | 18% | Tool descriptions, structured error envelopes |
| Reliability | 15% | Resume vs summary, retry logic, Batch API |
The traps that decide pass or fail
Three distinctions show up almost verbatim. Memorize which side each question is testing.
1. Programmatic enforcement beats prompt instruction
2. Improve tool descriptions before adding machinery
When an agent picks the wrong tool, the exam wants you to reach for the cheapest fix first:
Agent calls the wrong tool
├─ First: improve the tool description ✅
├─ Then: consolidate overlapping tools
└─ Last: add a routing classifier ❌ (usually the trap)Any option that jumps straight to a router or a classifier is almost always wrong.
3. Aggregate accuracy hides failure modes
When a question asks whether to reduce human review, an option that reports only aggregate accuracy should be eliminated. The honest metric is accuracy segmented by document type × field — that's what surfaces the failure mode hiding inside a good-looking average.
A four-week plan
- 1
Weeks 1–2 · Claude Code
Work Claude Code 101 then Claude Code in Action. Hooks, custom commands, and the Agent SDK are 20% of the exam and the fastest points to lock in.
- 2
Weeks 3–4 · Orchestration & tools
Do Introduction to Subagents and both MCP courses. Build one real MCP server that returns a structured error envelope — you'll never forget the pattern once you've written it.
- 3
Week 5 · Scenario drills
Run all six scenarios end to end. Keep an error log; every wrong answer is a study target.
Quick self-check
A support agent occasionally issues refunds it shouldn't. What's the exam-correct fix?
Study the weights, drill all six scenarios, and walk in knowing which side of each distinction the question is testing. That's the whole game.