CCAF Exam Prep — Claude Certified Architect (Foundations)

🇪🇸 ¿Prefieres español? Lee el README en español.

A hands-on study kit for the Claude Certified Architect – Foundations (CCAF) exam, mapped 1:1 to the official exam guide — every domain, every task statement, and all 6 official scenarios. The notebooks teach each concept by making it runnable: they make real Claude API calls (some via the Claude Agent SDK, a few via the raw Messages API purely for learning) so you watch the behavior the exam tests actually happen, instead of reading about it.

Important

Unofficial & independent. This repository is not affiliated with, endorsed by, or produced by Anthropic. It is a personal study project. I built it while preparing for the exam — I have not taken the exam yet (preparing as of 2026). Nothing here is a brain dump: everything traces back to the official exam guide and the official Anthropic Academy courses. The official guide itself is not redistributed here — download it from Anthropic (see The exam guide).

Why this repo

There are practice sites and repos that invent questions and drift away from what the exam actually covers. This one is the opposite: it is deliberately faithful to the official material. Each notebook section quotes the guide's task statement verbatim, then turns it into code you can run; each sample question is the official one, mapped back to the task statement it tests. If it isn't in the guide or the courses, it isn't here.

Status & validation

I personally designed, reviewed, studied, and ran every cell of the domain notebooks that exist — this is hand-built study material, not AI output dumped into a repo. The ccaf-notebook skill helps generate notebooks to a fidelity standard, but a skill's output is not validated on its own — the human pass is what makes a section trustworthy. Anything not yet marked ✅ is still being checked; verify it against the official guide before you rely on it.

Legend: ✅ hand-validated (designed · reviewed · studied · run) · 🔧 in progress · 🧪 built/generated, not yet hand-reviewed · ⏳ not built yet

Domain notebook	Status		Hands-on exercise	Status
D1 · Agentic Architecture & Orchestration	✅		01 · support-agent	✅
D2 · Tool Design & MCP Integration	✅		02 · claude-code-team	🧪
D5 · Context Management & Reliability	✅		03 · extraction-pipeline	🧪
D3 · Claude Code Configuration & Workflows	✅		04 · multi-agent-research	🧪
D4 · Prompt Engineering & Structured Output	✅		05 · cicd-review	🧪

(Exercise 01 has had my hand-review/tuning pass; 02–05 currently run and are portable but haven't been hand-reviewed yet.)

Start here — prerequisites (do these first)

This kit assumes you've already done the four official Anthropic Academy courses. Do them first, take notes by hand, and complete their exercises — that's where the muscle memory comes from. Their code lives in claude-with-anthropic-api/.

Official course	Mostly maps to
Building with the Claude API	D4 Prompt Engineering & Structured Output
Claude Code in Action	D3 Claude Code Configuration & Workflows
Introduction to Model Context Protocol (MCP)	D2 Tool Design & MCP Integration
Introduction to Agent Skills	D3 Claude Code Configuration & Workflows

Courses are on Anthropic Academy (and mirrored at github.com/anthropics/courses). Everything — the courses and the exam — is in English, so this kit is in English too (with this Spanish README as a convenience).

The exam at a glance

• 60 questions, 120 minutes, multiple-choice (1 correct + 3 distractors).
• 4 of 6 scenarios are chosen at random for your sitting.
• Pass = 720 / 1000 scaled. No penalty for guessing → never leave a blank.
• Domain weights: D1 27% · D3 20% · D4 20% · D2 18% · D5 15%.

What's inside

ccaf-prep/
  notebooks/      one notebook per domain + the mock exam
    D1_agentic_loops.ipynb          ✅  D1 · Agentic Architecture & Orchestration (27%)
    D2_tool_design_mcp.ipynb        ✅  D2 · Tool Design & MCP Integration (18%)
    D5_context_reliability.ipynb    ✅  D5 · Context Management & Reliability (15%)
    D3_claude_code_config.ipynb     ✅  D3 · Claude Code Configuration & Workflows (20%)
    D4_prompt_structured_output.ipynb ✅  D4 · Prompt Engineering & Structured Output (20%)
    mock_exam_and_review.ipynb      coverage checklist + the 12 sample Qs by scenario
    README.md                       the study-order guide
  exercises/      five cross-domain, runnable hands-on builds
    01-support-agent/      02-claude-code-team/      03-extraction-pipeline/
    04-multi-agent-research/         05-cicd-review/
  MAPPING.md      abstract task statement → concrete file:line → sample question (the index)
  STUDY_PLAN.md   the full schedule and rationale
  reference/      where YOU drop your downloaded exam_guide.txt (git-ignored — see its README)
claude-with-anthropic-api/   the official Anthropic course code + exercises
.claude/skills/ccaf-notebook/   the skill used to generate the notebooks (see below)

Each domain notebook follows the same anatomy per task statement: verbatim guide quote → plain-English table → a runnable cell that makes it observable → the anti-patterns as code (the exam's wrong answers) → a pointer to the matching line in your own course/exercise code → a self-check. Each ends with that domain's official sample questions (answers hidden).

How to study (recommended order)

The order is dependency-driven so each exercise unlocks as soon as its domains are covered — not pure exam weight:

1. Notebooks: D1 → D2 → D5 → D3 → D4. Run every cell; do the hidden sample questions actively (cover the answer, justify why each distractor is wrong, then reveal).

2. Exercises, in waves (they're cross-domain — none is single-domain):

   After studying…	Do exercises	Domains used
   D1 + D2 + D5	Ex1 then Ex4	D1 + D2 + D5
   + D3	Ex2	D3 + D2
   + D4	Ex3 ∥ Ex5	D4 + D5 / D3 + D4

3. Consolidate in mock_exam_and_review.ipynb: tick coverage as you finish each domain, then do the scenario-grouped mock at the end.

4. Finally, sit the official Practice Exam (Skilljar) under timed conditions.

Which layer: raw Messages API vs Agent SDK (and why)

The exam tests architecture concepts, not one SDK's syntax — so the notebooks use whichever layer the concept actually lives at, not a global preference:

• Raw Messages API (client.messages.create) when the concept is a Messages-API primitive the SDK hides and you need to see it: agentic loop / stop_reason (D1.1), tool_use/tool_result, tool descriptions & tool_choice (D2.1/2.3), structured errors / isError (D2.2), structured output via tool_use + JSON schema (D4.3/4.4), Batch (D4.5), single-turn decisions (D5.1/5.2/5.5/5.6). (Raw-API tool use has two halves: Claude returns a tool_use block — that's "calling" the tool — and your code executes it and returns a tool_result. The API never runs your function; that loop is the protocol.)
• Agent SDK (query() + Task / AgentDefinition) when the concept is a runtime feature — doing it raw would be faking it: coordinator + subagents + Task (D1.2/1.3), PreToolUse/PostToolUse hooks and deterministic gates (D1.4/1.5), sessions / fork_session (D1.7), multi-agent error propagation & delegation (D5.3/5.4).
• Claude Code (config/CLI) when it's configuration, not code: CLAUDE.md, .claude/commands/, .claude/rules/, plan mode, headless -p (D2.4, all of D3).

Rule of thumb: Agent SDK / Claude Code is the predominant architecture vocabulary (most of D1, D3, and the multi-agent parts of D5); the raw API exposes the primitive underneath when the concept is that primitive. Each notebook's setup cell states which layer it uses and why; full detail in ccaf-prep/notebooks/README.md.

Keep ccaf-prep/MAPPING.md open as your index. The full rationale is in ccaf-prep/STUDY_PLAN.md.

Setup

Requires Python 3.12+ and uv.

# 1. install dependencies (creates ccaf-prep/.venv)
cd ccaf-prep && uv sync

# 2. add your API key — copy the template at the repo root and edit it
cp ../.env.example ../.env
#    then put your key in ../.env:  ANTHROPIC_API_KEY=sk-ant-...

The .env is auto-discovered from anywhere in the repo (no path config needed). Calls default to claude-haiku-4-5 — the cheapest model — because these demos teach the mechanism, not model IQ. Override everywhere with CLAUDE_MODEL=... in .env.

Run a notebook: open it in VS Code / Jupyter and select the ccaf-prep kernel, or run all cells headless. Run an exercise:

cd ccaf-prep/exercises/01-support-agent && uv run python agent.py

Each exercise folder has its own README.md with the build steps, the task-statement tags, and the anti-pattern toggles to try (flip a flag, watch the wrong approach fail).

Keep your notebook outputs local, not committed

Notebook outputs are stripped from commits automatically (clean diffs, no leaked API responses) while staying in your working copy. Enable it once per clone:

cd ccaf-prep && uv run nbstripout --install --attributes ../.gitattributes

Track your own progress

The mock exam holds a coverage checklist. To record your progress without committing it, make a personal copy (it's git-ignored):

cd ccaf-prep/notebooks && cp mock_exam_and_review.ipynb mock_exam_and_review.personal.ipynb

Keep handwritten-notes scans or any private notes in a git-ignored personal/ folder.

The exam guide

The official guide is not included (it's Anthropic's material). Download the Claude Certified Architect – Foundations Certification Exam Guide from Anthropic. If you want the notebook-generator skill to regenerate domains, save the guide's text as ccaf-prep/reference/exam_guide.txt (git-ignored) — the 12 sample questions reproduced in the notebooks are the ones Anthropic publishes as samples in that guide.

How the notebooks were built

The notebooks were generated and maintained with a Claude Code skill, .claude/skills/ccaf-notebook/, which enforces the fidelity rules above (verbatim quotes, real API calls, anchored file:line pointers). It's included as a reference if you want to extend or regenerate domains.

About me

Javier Criado Gómez — AI Engineer. LinkedIn: https://www.linkedin.com/in/javierjcriadogomez/ · questions, suggestions, and corrections welcome (open an issue or a PR). I'm sharing this to help others certify while I study for it myself — once I've sat the exam I'll update this note.

Contributing

Found something that drifts from the official material, or a notebook cell that doesn't run? Open an issue or PR. The bar is fidelity to the official guide and courses — corrections that pull it closer to the source are the most valuable.

Disclaimer

Independent, community study material. Not affiliated with or endorsed by Anthropic. "Claude" and "Anthropic" are trademarks of Anthropic. The official exam guide and courses are Anthropic's; they are referenced and linked here, not redistributed. Provided as-is, for educational use.