How to sit Claude Certified Developer – Foundations: the score, the weights and the wrong answers

The rules the strategy rests on

• 720 on a 100–1,000 scaled score. Not 72%, and it converts to no number of correct items, so do not plan by counting how many you can afford to miss.
• Total score only. The domain percentages on your report are informational; there is no bar to clear in each domain.
• Criterion-referenced. The cut score comes from a formal standard-setting study, not from how other candidates did that week.
• Multiple-choice and multiple-response. Each item states how many options to select. In the captured courseware items, twelve of seventeen offered four options and five offered three — never assume a shape.
• $125 an attempt, four attempts at most in a rolling 12 months, with waits of 14, 30 and 90 days after successive misses.

Where your hours go

Because only the total counts, the rational plan is to maximise total items correct, so weight your time by what each domain carries:

(The guide publishes weights; the item column is each weight multiplied by 53.)

An hour that lifts your accuracy ten points in Applications and Integration is worth about 1.8 items; the same hour in Eval, Testing, and Debugging is worth about 0.1. Both are real improvements, and one is worth about a dozen of the other.

Go below the domains, because this guide lets you. It weights 25 sub-objectives. Claude Application Design (8.6%) outweighs the whole Security and Safety domain; Software Engineering Foundations (7.4%) outweighs Tool Implementation (4.4%), the heaviest sub-objective in Tools and MCPs; Claude Hooks (1.0%) is about half an item. Applications and Integration is not "the API chapter": requirements, the systems life cycle, engineering foundations, application design and configuration management are all graded inside it.

The one exception to ordering by weight is a domain where you are near zero rather than weak. Security items are decidable once you know the failure modes and undecidable when you do not, so going from nothing to competent there is a bigger swing than 8.1% suggests.

Find out which domains are yours to fix before you start: Alloy's free diagnostic is 24 questions, about 20 minutes, weighted like the exam.

How you read an item

Evidence first, then the layer, then the options. Many items hand you an error, a trace or a configuration. Read it before the options and classify the defect — schema validation, message ordering, transport or authentication. Then name the layer it lives in — prompt, parser, schema, loop, tool or transport. Only then read the choices, and discard everything outside that class and layer. This reading order is what makes the trace items fast.

Use what the stem closes off. A trace item that tells you the schema is valid and the description is specific has eliminated two of three fixes before you read them.

Watch for the five wrong-answer shapes, in their developer form:

1. 1.The absolutist — "always", "never", "only". One counter-example and it is out.
2. 2.True, but not the question — a correct statement about stop_reason or is_error in an item about ordering. Re-read the final sentence.
3. 3.Right action, wrong layer — a prompt change for a parser defect, a schema change for a loop defect.
4. 4.The over-engineered one — an agent where one call works, a retrieval index where a cached prefix works, a stronger model where a corrected prompt works. Prefer the simplest option that fully meets the requirement.
5. 5.An answer to a different question — restate the question in your own words, then match.

Two developer-specific traps sit on top of those. A plausible-but-wrong API shape names real fields in a real order and fixes a different class of defect than the error names — a required array added to a schema when the failure is a mismatched tool_use_id. An anti-pattern wearing best-practice clothing promises thoroughness or convenience while removing a control: rotating a key but leaving it in the credentials file, concatenating every past transcript "for memory", or switching to bypassPermissions for fewer prompts.

Practise subtraction. One item shape hands you four plausible pieces of configuration and tells you to leave out anything that does not belong. That clause is the whole item; the over-engineering instinct loses it.

Change an answer only with a receipt. "I missed the word only." "I overlooked the constraint." "The question asks which action comes first." Those are auditable reasons. A wobble in confidence is not.

Pace for reading. Two minutes an item is comfortable for recall and tight for a trace. Nothing published says whether you can flag and revisit items, whether delivery is adaptive, whether partial credit or negative marking applies, or whether breaks and scratch paper are allowed; confirm anything your pacing depends on when you register.

How you practise

• Read each domain as failure prevention. Anthropic's courseware frames every topic as the implementation decision that prevents a specific production failure. Ask that question and scenario items stop looking like trivia.
• Review your reasoning, not the letter you picked. The right option for the wrong reason will not survive a rephrased trace.
• Rate your confidence honestly. Certain-and-wrong answers never get re-checked; Alloy collects them into a practice set of their own.
• Keep flashcards due. Reviewed just before you would forget them, exact values — status codes, field names, defaults — stay available without a cram session.
• Know the fundamentals precisely. The exam rewards distinctions such as choosing a model versus enabling extended thinking, streaming versus async, max_tokens as the only hard ceiling, and why temperature: 0 does not make output identical — the material of the course's free second lesson.

When you sit, and what you do after

The binding constraint is not the retake ladder but the four-attempt annual cap, with the full fee each time. Book when your practice says you are ready, not when a date feels motivating. In Alloy that is the readiness figure, built from mastery, honest confidence and coverage, and adjusted — within bounds — by practice exams.

On the day: accept the NDA (declining ends the session with no refund), bring government photo ID matching your registration exactly, and never cancel inside 24 hours. After: keep the per-domain breakdown if you need a retake — it is your plan. Appeals run within 14 days but cannot contest the standard-setting outcome or item content. Renew within 12 months and it is a free, non-proctored assessment; let it lapse and you sit the full exam again.

Next step

Take the free CCDV-F diagnostic, then the free exam-strategy lesson and its podcast. The exam-strategy lesson is worth a 10–20% score bump.