How to answer Claude Certified Architect questions

Every CCAR-F question is printed in the same order: a scenario, a symptom, sometimes a constraint, and a final sentence that asks for something. That is the order it was written in, and the wrong order to read it in. The scenario comes first on the page and matters least. The last sentence decides which kind of answer wins.

Read the question backwards

1. 1The ask, the final sentence. It names the kind of answer that wins.
2. 2The constraint, if there is one: a phrase like must never, every time or cannot be bypassed. It rules out whole families of answers at a stroke.
3. 3The symptom, the observable failure. It points to the topic being tested.
4. 4The scenario, the production setting. It tells you the domain and decides little.

Know what the ask wants

• The root causeWhat winsThe earliest point where information was lost or never assignedWhat losesEverything downstream of it, however true
• The best first stepWhat winsThe lowest-effort change that addresses the causeWhat losesValid fixes that cost more than the first step needs
• The most effective fixWhat winsThe highest-leverage fix, whatever it costsWhat losesPartial fixes that leave the cause in place
• The correct approachWhat winsThe mechanism whose guarantee matches the stakesWhat losesAnything probabilistic when the question says always
• One answer per itemWhat winsOne rule applied the same way to every itemWhat losesOptions that treat every item alike

Root cause and best first step are the pair to watch. They are asked about identical scenarios and have different right answers. An answer that is correct but expensive can be the right root-cause fix and the wrong first step.

Cross out in this order

Remove what cannot be right from the four, in a fixed order. The first three steps are meant to leave you with two options.

1. 1Cross out anything that catches the problem after it happens when the question demands prevention.
2. 2Cross out anything that only adds more: more retries, a bigger context window, a higher-tier model.
3. 3Cross out any setting or flag you have never come across. Invented mechanisms are planted as distractors.
4. 4Re-read the final sentence and decide whether it asks for the root cause or the first step.
5. 5Between the last two, choose the fix for the cause over the adjustment to a number.

When two options are both true

Ending up with two defensible options is the normal finishing position, and it is deliberate. Break the tie with these, in order:

1. 1Deterministic over probabilistic when the question names a guarantee. A prompt is followed most of the time; code runs every time.
2. 2Cause over symptom. Fix the shape of what comes back instead of how many times you retry it.
3. 3The cheapest sufficient fix when the ask says first.
4. 4Existing over custom: a community MCP server, a built-in tool or the current model before anything you would build.
5. 5Scoped over global: a rule scoped to a path over the root config, one tool over full access.
6. 6Partial results with metadata over silence or a hard stop.

Read the numbers in the question

Figures in a question carry information. A tool count, an accuracy rate or a time limit points at one specific rule, and it can rule out two options on its own. A long list of tools points at tool design. An accuracy figure that looks reassuring across the whole dataset points at measuring by category.

Keep to the clock

60 questions in 120 minutes is about two minutes each. A wrong answer costs nothing a blank one does not, so never leave one empty. If a question is eating your time, eliminate what you can, choose between what is left and move on.

Technique gets you from four options to two. Knowing the mechanisms gets you from two to one, and that part is retrieval practice. Claude Certified Architect wrong-answer patterns and Never the answer on the Claude Architect exam are the fastest place to start.