~/peterdongo.dev

exams

Four Claude certification practice exams.

Anthropic runs a small family of Claude certifications - an Associate track for business users, a Developer track for people building on the API, and a two-level Architect track for designing systems around the model. While preparing for them I wrote four full-length practice exams, and I host them here with the tips and materials that actually helped.

Original · unofficial · free · runs entirely in your browser

§ 02
how they work

Multiple choice and multiple response

Every item tells you how many answers to select. Multiple-response items only score when you pick exactly the right set.

Scaled scoring, pass at 720

Your raw score maps to an approximate 100-1000 scale with a percent-correct breakdown per domain - just like the real score report.

Timed, 120 minutes

Generous for the item counts, but the scenario-based Architect questions eat clock. Read carefully.

Blueprint-weighted

Questions are distributed across the published domain weights, so a weak domain shows up clearly in your results.

§ 03
tips

You're building a reliable system around a fallible model.

Almost every right answer follows from that one sentence. These six are the rest.

01

Constrain, don’t add.

“More tools / more agents / a bigger context window” is almost always the wrong answer. The right fix usually removes a degree of freedom rather than adding one.

02

Prompts suggest; systems enforce.

If a rule must always hold - money, compliance, ordering - enforce it in code or a hook, not by asking the model nicely.

03

Fix the earliest layer.

Most wrong answers are reasonable fixes applied too late in the pipeline. Trace the failure back to where it first became possible and fix it there.

04

Match the fix to the failure.

Vague rule → explicit criteria. Inconsistent output → few-shot examples. Invalid JSON → a schema. Valid-but-wrong output → semantic validation.

05

Don’t collapse states.

“Tool failed” is not “found nothing.” Escalate on a policy or workflow trigger - never on sentiment or the model’s stated confidence.

06

Remember what you’re building.

You’re building a reliable system around a fallible model. So “just fine-tune it” or “tell it to be more careful” are almost never the answer.

§ 04
how I prepared

What I actually leaned on.

Beyond these practice exams. The other mocks below are external - genuinely useful for getting a feel for the question style.

These practice exams are unofficial and not affiliated with, endorsed by, or produced by Anthropic. They were generated with Claude as study aids and written against the published exam guides - not drawn from any live question bank. “Claude” and “Anthropic” are trademarks of Anthropic.