I passed Claude Certified Architect Foundations (CCAF) with 910/1000 on 15 July 2026, and I'm the first at my company to finish it. I open-sourced the study notes I built: ** github.com/ahmdhsn-dev/ccaf-study-guide**.
CCAF is scenario-heavy: you pick the right architecture when an agent forgets context, a generic tool keeps misfiring, or a customer escalation lands mid-process. The official guide covers all of it, but it's dense. The repo distills 30 exam tasks across 5 domains into two complementary views: what to do (patterns and traps) and what not to do (anti-patterns).
What's inside: a one-page concept map, a trap taxonomy (why wrong answers sound right), and an anti-patterns catalog that maps every "avoid / do not / instead of" warning from the official guide. If a distractor sounds plausible, check the traps first; if it sounds like a shortcut the guide rejects, check the anti-patterns.
Use it as your last cheat sheet, not your textbook. Take a mock first to see where you stand, then study the official guide and use the repo to close the gaps.
Full workflow: study/README.md
Building with the Claude API → Claude Code in Action → Introduction to Agent Skills → Introduction to MCP
I completed all four with hands-on labs in May 2026. Production work with skills, MCP servers, and tool orchestration helped, but the exam tests guide patterns, not how your repo is wired.
| Step | Doc | What you get |
|---|---|---|
| 1 | ||
Domain weights: Agentic Architecture (27%) · Tool Design & MCP (18%) · Claude Code Config (20%) · Prompt & Structured Output (20%) · Context & Reliability (15%). See the concept map for task breakdown.
Take the CyberSkill mock first (closest to the real exam), then the Certification Guide short mock. My first CyberSkill run scored 850/1000; after mapping misses to the guide and retaking, I was consistently 950+ before booking. For every miss, map it to the trap taxonomy, cross-check the anti-patterns catalog, re-read the matching section in the official guide, and repeat until you're consistently at 900+ (I booked at 950+).
Mocks are your prerequisite gate, not an arbitrary iteration cap. You won't feel ready. The exam has stop_reason: end_turn
before you do, and that's normal.
Nine misses on my first CyberSkill run; these three patterns accounted for most of them (full mapping in the cheat sheet below):
analyze_document
with an enum modeHigh-yield patterns from mock exams. Full taxonomy: 02-mock-exam-trap-guide.md.
Context filling up (multi-issue) → 5.1 Summarize resolved; active thread full
Generic tool, inconsistent output → 2.1 Split typed tools
Messy source formats → 4.3 Schema + normalization in prompt
Function behind wrappers → 2.5 List aliases → grep each
Agent forgets mid-conversation → 5.1 Pass full messages array
Unknown bug path → 1.6 Adaptive investigation
Long code exploration (30+ min) → 5.4 Scratchpad file
Mid-process escalation → 1.4 Structured handoff brief
Angry + no context → 5.2 Acknowledge + one question
Identity multi-match → 5.2 Ask identifiers, never guess
| Trap | Sounds like... | Guide-preferred |
|---|---|---|
| Sounds enterprise | Structured data layer, reference IDs | Progressive summarization |
| Sounds efficient | Re-fetch via MCP, parallel workers | Trim/summarize context |
| Sounds smart | Confidence threshold, routing classifier | Explicit criteria; programmatic gates |
| Sounds helpful | Immediate escalation, unauthorized retry | Calibrated: brief, one question, or resolve |
| Sounds thorough | Plan everything first | Adaptive evidence-driven steps |
| Sounds simple | Better description on generic tool | Split tools + typed schemas |
| Sounds pragmatic | Post-processing normalization | Canonical format at extraction |
From 03-anti-patterns-catalog.md:
1. Heuristic customer selection on multi-match: ask for email, phone, or order ID; never confidence-threshold auto-pick (Task 5.2).
2. Sliding window drops resolved threads: progressive summarization on resolved issues; keep active thread full (Task 5.1).
3. Generic analyze_document + free-text mode: split into
extract_data_points
, summarize_content
, etc. (Task 2.1).4. Batch API for pre-merge blocking checks: sync Messages API for blocking; Batch for overnight audits (Task 4.5).
5. Giving synthesis agent web search tools: scope to verify_fact
; route new lookups through coordinator (Task 2.3).
| Topic | Exam (guide) | Production (2026) |
|---|---|---|
| Structured output | ||
tool_use + JSON schema |
||
Also output_config.format ; prefill removed on new models |
||
| Batch tool use | No multi-turn in one request | Server tools + _turn ; client MCP loops still multi-request |
/memory |
||
| Diagnostic command (in scope) | Plus expanded CLAUDE.md / rules / skills ecosystem |
Full breakdown: 04-api-and-product-faq.md
** github.com/ahmdhsn-dev/ccaf-study-guide**. MIT licensed, unofficial, not affiliated with Anthropic.
Community links and extra mocks: resources/links.md
Disclaimer: Based on the exam taken 15 July 2026. Unofficial, not affiliated with Anthropic. Verify latest exam requirements before you sit.