How to Pass CCAR-P

Passing comes down to three things, and only one of them is knowledge. You need coverage across seven weighted domains, you need a practice method that changes how you read a question rather than just adding facts, and you need a clock plan, because 63 questions inside 120 minutes leaves under two minutes an item and is the tightest ratio of the four exams on this site.

The standard is a scaled score of 720 on a range running from 100 to 1000. It is criterion-referenced, meaning you are measured against a fixed bar rather than against the other people sitting that week. No conversion between raw answers and scaled points is published, so any target expressed as a percentage is your own invention, and planning around one is a mistake this page will come back to.

What follows assumes you already work on production AI systems and are not starting from zero, which is what the exam assumes too. The plan has three phases: map your position honestly, close the gaps the map exposes, then rehearse under the real clock until the pace stops being the thing you notice.

A note on register. This paper rewards restraint. Across every domain, the option that displays the most knowledge is frequently the distractor, and the correct answer is the plainer thing that survives the constraint written into the stem. Preparation that does not train that instinct will leave you well informed and still short of the line.

Phase one: map before you study

Sit a full-length timed run before you open any material. Cold, no notes, the real count and the real clock. It will feel wasteful and it is the highest-value two hours in the entire plan.

The output you want is not a score. It is a per-domain breakdown, so that your preparation is aimed at measured weakness rather than at the topics you find interesting. Almost everyone discovers at least one surprise, and it is usually in the same place: the governance, stakeholder and lifecycle material that technical candidates assume they can reason through on the day.

Write the result as a short table: domain, blueprint weight, your accuracy. Multiply the gap by the weight and you have your priority order in a form that is hard to argue with. A domain worth 14 percent where you scored badly outranks one worth 7 percent where you scored slightly worse.

Do this again roughly two-thirds of the way through preparation. Plans built on a single early diagnostic go stale, because the drilling works and the weak spots move. Two data points also tell you something one cannot: whether you are actually improving, or simply becoming familiar with the questions you have already seen.

Phase two: close gaps by writing, not by rereading

Rereading produces recognition, and recognition does not survive four plausible options under time pressure. Everything in this phase should produce something written.

For each priority domain, work through questions in blocks and treat the explanation as the actual material. Read the rationale for every option including the ones you correctly rejected, because the reason a wrong answer is tempting is the pattern being tested. Then write one line per missed item: what the stem constrained, and what the attractive wrong answer would have cost in production. Cost is the operative word. If you cannot name a consequence, you have not understood the item well enough to get its cousin right next month.

Keep a running list of discriminators rather than a list of facts. Whether subtask count is known in advance. Whether an action is reversible. Whether a control sits inside the model's reach or outside it. Whether a commitment is backed by measurement. These pairs are what the questions turn on, and a page of them is worth more in the final week than any set of notes.

When a domain stops producing new lines for your list, it is done for now. Move to the next one on the priority order.

What a scaled 720 means for how you play the paper

The scale runs 100 to 1000 and the bar is 720, but that is not 72 percent of anything. Because no raw-to-scaled conversion is published, you cannot compute how many items you are allowed to miss, and every plan that starts with a permitted number of wrong answers is built on a number nobody has released.

Three consequences follow, and they are all practical.

First, no domain is disposable. Without a conversion table you cannot show that the smallest domain is safe to skip, and on a paper this size a handful of items decides borderline results. Coverage beats depth in one favourite area.

Second, leave nothing blank. Nothing published suggests a deduction for a wrong answer, while an unanswered item is guaranteed to earn nothing. Where an item asks you to pick more than one response, partial knowledge is worth less than it feels: a selection that is half right is not half a mark, so read the instruction about how many to choose and honour it exactly.

Third, ignore any comparison with other candidates. A criterion-referenced standard is fixed in advance, so there is no curve to be rescued by and no benefit in wondering whether the paper was hard for everyone. Measure yourself against your own timed runs and against consistency, which is the only signal that transfers.

Pacing: 63 items, 120 minutes, no slack

Divide it out and the average is one minute and fifty-four seconds per question, with nothing left over for review. This is the only one of the four Claude papers here that gives you under two minutes an item, so pacing deserves more deliberate practice than it would elsewhere.

Aim to average about one minute forty-five instead. That banks roughly ten minutes and it is achievable, because a real paper mixes short items in with the long ones. Set four checkpoints and glance at the clock only at those: 17 questions by the 30-minute mark, 34 by 60 minutes, 51 by 90 minutes, and the last item answered around 110. Checking after every question costs attention and buys nothing.

The discipline that makes those checkpoints hold is a hard rule about the second read. Read the stem, find the constraint, eliminate against it. If two options still stand after that, choose the plainer one, flag it, and move immediately. Deliberating for four minutes between two defensible designs is the specific way experienced architects run out of clock, and the extra time rarely changes the answer.

If you fall behind, recover by shortening your deliberation, never by skipping items. Skipped questions become end-of-paper panic and get answered worse than they would have been in sequence.

Phase three: rehearse until the clock stops registering

In the final stretch, replace study blocks with full simulations at the true count and duration, and stop introducing new material about five days out.

What you are training is not recall, it is composure. The first full run under a genuine clock feels bad for almost everyone. The third one feels like a task. Getting that adaptation done in practice, rather than on the day, is worth more than another pass through your notes. Sit at least one simulation at the same hour your appointment is booked for, since a paper this dense is a different experience at eight in the morning than at three in the afternoon.

Between runs, review only what you got wrong and only through the explanations, adding to your discriminator list rather than starting new topics.

The readiness signal is stability rather than any single result. Two or three consecutive timed runs at a comparable standard, with no domain sitting far below the rest, is the point at which booking makes sense. A single strong score after several weak ones is variance, and paying a full fee on the strength of it is how people end up in the retake ladder.

Sort logistics in the same week too, since a booking can only be changed up to 24 hours ahead of the appointment itself.

If it goes wrong: the retake ladder

A failure is recoverable and the terms are worth knowing before you book rather than after. The wait is 14 days before a second attempt, 30 before a third, and 90 before a fourth, with no more than four attempts inside any rolling twelve months. Each sitting is charged in full.

The waits are long enough to matter to a plan. Two failures put roughly six weeks between you and a third attempt, which is why a booking made on optimism rather than on stable practice scores is expensive in time as well as money.

If you do fail, resist the urge to restudy everything. The score report is the most precise feedback you will get, and where it breaks results down by domain it is precise enough to build the next attempt on. Take the two weakest weighted areas, work through them with the writing method above, and only then return to full simulations. A second attempt spent rereading material you already knew tends to produce a similar score.

Use the waiting period rather than resenting it. Fourteen days is enough to genuinely close one domain. Ninety is enough to build the production experience the paper is really testing, which for some candidates is the more useful outcome. Questions about your result or the process go to certifications-support@anthropic.com.

Common questions