CCAR-P - Solution Design & Architecture (17% of the exam) - Section 1.5

Apply decomposition techniques for complex problem solving.

Breaking a complex task into smaller steps that can each be prompted, validated and retried on their own, such as prompt chaining, routing to specialised handlers, and parallel sub-tasks with a merge step. Candidates should choose a decomposition that makes each step checkable.

prompt chainingroutingparallelisation and mergeindependently verifiable steps

Practice question for this objective

Free sampleSolution Design & Architecturemedium

A software company's support assistant drafts email replies through a seven-call prompt chain: detect the language, extract the product and version, retrieve knowledge-base articles, write an outline, expand the outline into a reply, adjust the tone, and check that every claim cites a retrieved article. Code validates the extracted product and version against the release catalogue, and legal requires the citation check to pass before any reply is sent. The outline, expansion and tone calls each rewrite the previous call's prose with no check between them, and together they take 11 of the chain's 14 seconds at p95 against a 7-second target. In trials, a single drafting call matched their combined rubric score in about 3 seconds, while moving every call to a smaller model tier lowered the rubric score from 4.4 to 3.7. What should the architect recommend?

ARun all seven calls in parallel on the incoming email and combine their outputs in code before the reply is sent
BDrop the citation check call and rely on the drafting prompts' instruction to cite only retrieved articles
CMerge the outline, expansion and tone calls into one drafting call, keeping extraction and the citation check separate Correct
DMove every call in the chain to the smaller model tier and add a second tone pass to recover the lost rubric score

Split a chain at points where a step's output can be independently verified, and merge adjacent steps that add latency without adding a checkable boundary. Decomposition pays for itself when each step produces an output that can be validated, retried or gated on its own. Here the extraction step is validated in code and the citation check is a required gate, so both boundaries carry value. The outline, expansion and tone calls only pass prose from one to the next with no check, so they add round trips without adding verifiability; collapsing them into one drafting call recovers most of the latency budget and keeps every checkpoint the business depends on.

Why A is wrong: Parallelisation is tempting because it attacks wall-clock time directly, and it is the right pattern for independent subtasks. These steps are not independent: retrieval needs the extracted product, expansion needs the outline and the citation check needs the finished reply, so they cannot run side by side.

Why B is wrong: Removing a call does cut latency, and an instruction to cite sources often improves grounding. It trades away the one step legal requires to pass before sending, replacing a verifiable gate with a prompt instruction that nothing checks.

Why C is correct: A chain boundary earns its latency when the step's output can be checked on its own. The three prose-rewriting calls have no check between them, so merging them loses no verification point, matches their quality in the trial and brings p95 to about 6 seconds, while the validated extraction and the required citation gate stay as distinct, checkable steps.

Why D is wrong: A smaller tier is usually faster, which makes this a natural first idea for a latency target. The trial already showed the quality loss, and adding another rewriting call adds latency and another unchecked step rather than removing the steps that carry no verification value.

See more CCAR-P practice questions, answers explained.

Exam traps in Solution Design & Architecture

Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-P bank for this domain.

Move both steps to the smaller tier, since its extraction results show it is adequate for this agency's repair requests
Why it is wrong: The extraction result is real evidence and moving everything gives the largest spend cut, so this is a natural reading of the trial. It generalises one step's evaluation to a different step: the smaller tier scored 88 percent on urgency banding, below the 96 percent the agency must hold.
The technical handler's model has degraded since the release, so the team should move that handler to a larger model tier.
Why it is wrong: Tempting because a falling resolution rate looks like a model regression. But the handler's model is unchanged and its own evaluation set still scores 88 percent when run directly, which rules out the handler itself as the source of the drop.
Keep the single prompt, move it to a larger model tier, and add an instruction to re-check every medicine against the pharmacy extract before drafting.
Why it is wrong: A more capable model and a self-check instruction are tempting because the error looks like a reasoning slip. It is wrong because it leaves all three tasks fused in one output, so failures stay unattributable, and it gives the pharmacist no reconciled list to approve, which the governance requirement demands.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.