How to pass the Claude Certified Architect – Foundations exam

Most advice about this exam is a restatement of the blueprint. The blueprint is already public and summarised here, so this is about the part that is not written down: what the exam rewards, where people lose marks they did not need to lose, and what order to do things in.

The exam is a judgement test, not a recall test

This is the single most useful thing to internalise, and it changes how you should study.

A recall test asks what a thing is. A judgement test gives you a system that is already misbehaving and four defensible-sounding responses, and asks which one a competent architect would pick. Every one of the four is something a real person has actually tried. Three of them address a symptom.

Consider the shape of a typical item. An extraction pipeline is producing inconsistent output. Do you (a) add more explicit instructions to the prompt, (b) lower the temperature, (c) introduce a confidence threshold and discard low-confidence results, or (d) add a few-shot example covering the ambiguous case?

Every option is a real technique. Three of them are things engineers reach for first. Only one is what the documentation actually recommends for this failure mode. Knowing the four techniques exist gets you nothing here — the marks are in knowing which one the situation calls for.

This is why people with genuine production experience still fail. Experience gives you a strong prior about what you would do. The exam is testing whether you know what the guide says the right answer is, and those diverge more often than is comfortable.

Where the marks actually are

The five domains are not weighted evenly, and the spread is wide enough to matter:

Domain 1 is worth nearly twice Domain 5. It is also the domain whose concepts recur inside the other four — agentic loops, orchestration patterns and context passing show up in the tool design questions and the reliability questions too. Studying it first compounds.

The scoring makes this sharper than the table suggests. You pass at a scaled score of 720 out of 1,000, and scaling means a heavily weighted domain moves your total more than a light one. Being strong everywhere except Domain 1 is a worse position than being merely adequate everywhere.

The scenarios are published — use that

The exam frames every question inside one of six scenarios, four of which appear in any given sitting. All six are published in the exam guide.

That is unusually generous, and most candidates under-use it. Read all six before you study anything. They tell you what kind of system the exam thinks is worth asking about — a customer support agent with a refund tool, a multi-agent research pipeline, Claude Code inside CI — and once you have read them, the abstract domain descriptions stop being abstract.

It also means you can practise recognition. When an item opens with an agent that has process_refund in its toolset, you should already know you are in scenario 1, that escalation criteria and tool-description quality are what it cares about, and that the answer is unlikely to be "add it to the prompt."

What actually goes wrong

Four failure patterns, in roughly the order they cost people marks.

Answering from this week's documentation. The exam guide is version 1.0, effective July 2026. Anthropic ships faster than that. Where a product has moved on since the guide was written, the exam still marks against the guide. This is a genuinely uncomfortable rule and it is the right one to follow — the divergence log exists precisely to record the cases where the two disagree.

Reading instead of building. The judgement the exam tests comes from having made a choice badly once. You can read about PreToolUse hooks and retain nothing useful; you cannot write one that fires at the wrong moment and forget what you learned. Every lesson here has a build exercise for that reason, and skipping them is the most common way to over-prepare and under-perform.

Confusing an iteration cap with a termination condition. Not really one mistake — a family of them. The exam repeatedly tests whether you understand the mechanism rather than the workaround. A loop ends because stop_reason says so; a cap is a safety bound. Context degradation is not running out of window. A confidence score is not calibrated because it looks confident. In each case there is a shallow answer that sounds right and a mechanical one that is right.

Over-preparing on out-of-scope material. The guide publishes a list of what will not appear: fine-tuning, model internals, Constitutional AI and RLHF, embeddings and vector databases, computer use, vision, streaming, rate limits and pricing, OAuth and key rotation, cloud provider configuration, MCP server hosting. If you find yourself deep in any of those, stop. That time belongs to Domain 1.

A preparation order that works

Sit a mock exam cold, before reading anything. A score you have not studied for is the only honest map of your gaps, and it costs two hours you would otherwise have spent re-reading things you already know. Expect it to feel bad. That is the point.

Read the six scenarios. Fifteen minutes, and it reframes everything after it.

Work the domains in weight order — 1, then 3 and 4, then 2, then 5. Within each domain, work the task statements in order; they build.

Do the build exercises. At minimum: one complete agentic loop with tool calls and session resumption, one properly configured repository with a CLAUDE.md hierarchy and a path-specific rule, one MCP server with deliberately similar tool descriptions, and one extraction pipeline with schema validation and a retry loop. That is the guide's own recommended preparation, and it is recommended for a reason.

Re-test and watch the per-domain numbers, not the total. The overall percentage hides the one domain that is going to cost you the pass. Keep going until the weakest domain moves.

The night before, read the cheat sheets and the divergence log. Not new material. The cheat sheets are one dense page per domain, built for exactly this.

How long it takes

The guide describes an ideal candidate with around six months of hands-on experience across the Claude API, Agent SDK, Claude Code and MCP. With that background, 15–20 hours of focused preparation is realistic. Without it, budget 30–40 and put the extra hours into building rather than reading.

If you have never written an agentic loop, that gap is not something a study guide closes. Write one first, then come back.

One thing worth being sceptical about

Any practice score, including the one on this site. The real exam reports a scaled score, and Anthropic has not published the scaling function — nobody outside the programme can reproduce it. A 690 on a mock does not mean you are 30 points short.

Use practice scores for their one honest purpose: finding which domain to study next. The mock exam reports per-domain accuracy underneath the headline number, and that breakdown is the part worth acting on.