I hadn't sat a technical certification in a long time. I didn't need one for my CV. So why go all in on the Claude Certified Architect? Because I believe what's coming is big enough that I wanted to earn it myself, hands on keyboard — not delegate it, not talk about it from the outside. If I'm going to ask Malaysians to do this, I should do it first.
"Giving back to society matters — especially the goal of making Malaysia an AI nation by 2030. To get there, we have to move from being AI consumers to AI creators. The Claude Certified Architect is genuinely step number one."
This guide isn't theory. I sat the Claude Certified Architect — Foundations exam on 3 June 2026 and passed. Here's the real scorecard so you know exactly what "ready" looks like.
Most of us grew up with one exam strategy. JPJ undang? Drill the 500-question bank. KPP? Memorise the answers, not the road rules. SPM sejarah? Hafal the essay format, plug in dates, collect marks. None of that works here.
Anthropic designed CCA-F so the practice and real exams test the same concepts — but almost none of the actual questions match. You can score 980 on the mock and still fail if you memorised answers without the reasoning.
This isn't about what Claude can do. It's about making sound architectural decisions that hold up when production goes wrong. You can't bluff. You have to actually know.
What I felt opening question one. A production support system behaving incorrectly. 180 words of context. Four fixes — all reasonable, two really reasonable. The wrong answers aren't obviously wrong; Anthropic calls them distractors, built to catch people who get the concept at the surface but never internalised the why. I finished in under an hour — but I'd spent weeks building that confidence. Not memorising. Building.
Not how many hours you studied. Not whether you finished every course. These six.
The single most pervasive concept on the exam. Prompts are guidance. Code is law. "Verify identity before refunding" works 95% of the time — and in production that 5% is unacceptable.
"enhance the system prompt to…" is almost always wrong when money, compliance, or a hard business rule is involved.
Multi-agent is the biggest single domain. A spawned subagent has absolutely no idea what came before — not the user's request, not prior findings, not what tools ran. A blank page, not "limited context."
Everything it needs — prior results, document content, output format, quality criteria — must be written into its prompt explicitly, every single time.
Claude doesn't pick tools by name — it reads descriptions. Two thin descriptions = constant misrouting. A good description states exactly what it returns, accepted inputs with examples, when to use it vs. an alternative, and edge cases.
Read-only data is not a Tool. If Claude just needs to know something, that's an MCP Resource. Tools are for actions. Getting this boundary wrong is the most common source of lost marks.
Claude attends more reliably to the start and end of its window. A constraint buried at position 80,000 of 200,000 is processed less reliably than the same text up top — the "lost in the middle" effect.
Fix by restructuring where information lives: persistent "case facts" blocks, key findings at the top of aggregated inputs, scratchpad files for codebase exploration — not just a bigger window.
Token economics has a table you should know cold. The trap: "switch everything to Batch to save 50%." Any workflow where someone is waiting right now — pre-merge checks, live support, interactive review — can't use Batch's 24-hour window.
| API | Cost | Latency | Use when |
|---|---|---|---|
| Real-Time | Standard | Immediate | User is actively waiting |
| Prompt Caching | Up to 90% off | Immediate | Same context repeated |
| Message Batches | 50% off | ≤ 24 hours | Background, non-blocking |
Repeated context + real-time need → Prompt Caching, never Batch.
The meta-skill. By exam day you should flag a distractor as wrong on pattern recognition — before you've read the right option.
Each presents a realistic production context that frames a cluster of questions. You'll get 4 of these 6 — you don't know which.
Support agent with lookup, order, refund, and escalation tools. Tests escalation logic, tool-sequencing enforcement, multi-issue handling.
Team dev workflow. Tests CLAUDE.md hierarchy, path-specific rules, plan mode vs. direct execution, Explore subagent usage.
Coordinator dispatching to research, analysis, synthesis, report subagents. Tests explicit context passing, parallel spawning, error propagation with coverage gaps.
Agent SDK to explore codebases and automate tasks. Tests built-in tool selection (Grep vs. Glob vs. Read/Edit/Write), MCP resource design, scratchpad persistence.
Integrating Claude Code into pipelines for review and test generation. Tests the -p flag, JSON output flags, session isolation for independent review instances.
Extracting structured data from documents at scale. Tests tool_choice, nullable schema fields, validation-retry loops, batch strategy, field-level confidence routing.
The official FAQ is blunt: "Without hands-on experience with Claude Code, the Agent SDK, the Claude API, and MCP — skipping this step will likely result in you not passing."
All four Academy courses. Build vocabulary cold.
The most important week. Hands on keyboard.
All 6 scenario labs. Practice exam (target 900+). Find gaps.
Anti-pattern drilling. Capstone build. Final readiness check.
Specific syntax and settings. These show up as exact-recall questions.
Don't study these for CCA-F purposes:
I've seen people go in unprepared who are now waiting out their 6-month retry window. The exam costs $99. It's proctored. One attempt. That's not designed to be punishing — it's designed to mean something. A CCA badge should signal you can architect Claude systems at a professional level, and that value only holds if the bar stays high. Go in ready. The preparation is the point — not the certificate.
Foundations proves you can architect Claude systems. But Anthropic has signalled what comes next: a Forward Deployed Engineer (FDE) Black Belt level. The people who pass now are first in line.
The credential you can earn today. Sound architectural judgement across the API, Claude Code, Agent SDK, and MCP. This guide gets you there.
The earlier you pass Foundations, the earlier you're positioned for what Anthropic builds on top of it. The window to be early in Malaysia is open right now.
First-hand prep tips, discussion on the concepts the exam actually tests, and — once you pass — a place among the first Malaysians to hold the badge. Completely free to join, leave anytime.
Join the Free WhatsApp Community →Want more than self-study? Build under guidance, ask questions live, and get tested before the exam. Two focused days take you from fundamentals to exam-ready.
Seats are limited. Join the free WhatsApp community to be first to know when registration opens.
Enquire about the 2-day intensive →Our target of 1,000 Claude Certified Architects from Malaysia by 2026 isn't about numbers. It's about making sure that when enterprises here decide how to build AI, there are Malaysians in the room who know what they're doing — who can speak to architecture, not just features; who can name what will fail in production, not just what works in a demo. You can be part of that. Start today.
The Claude Certified Architect — Foundations is Anthropic's certification that tests whether you can make sound architectural decisions when building production systems with Claude — across the Claude API, Claude Code, the Agent SDK, multi-agent systems, and MCP. It's 60 multiple-choice questions in 2 hours, scored 100–1,000, with a passing score of 720.
A scaled score of 720 out of 1,000. For reference, AK — who wrote this guide — scored 822 (48 of 60 correct, 80% raw accuracy) and passed.
The exam costs USD $99. It's fully proctored, you get one attempt, and if you fail you wait 6 months before retaking. Go in ready.
Don't memorise practice answers — build. Finish the four free Anthropic Academy courses, then build five hands-on projects (validation-retry loop, multi-agent with explicit context passing, a CLAUDE.md hierarchy, a CI/CD simulation, and an MCP server with a Tool and a Resource). Budget ~30–37 hours over about four weeks. The full path is above.
Yes — it's 100% free. It exists to share prep tips, discuss the concepts the exam tests, and connect Malaysians working toward the certification. Join here.
Foundations is the entry level. Anthropic has signalled an upcoming Forward Deployed Engineer (FDE) Black Belt certification — and those who pass Foundations may be among the first future Black Belt FDEs and Claude Champions. Passing now positions you early.
Claude is a registered trademark of Anthropic. Agmo is an independent service provider and training company. The Claude Certified Architect (CCA-F) is a certification programme operated by Anthropic; Agmo's architect team members hold this credential and provide preparation services referencing it. This study guide reflects one architect's experience and is not affiliated with or endorsed by Anthropic.
Proudly designed by Claude Design and coded by Claude Code with ❤️
