CCAR-P - Claude Models, Prompting & Context Engineering (13% of the exam) - Section 2.5

Implement prompt reuse strategies (caching, modular prompts, Skills).

Reusing prompt content efficiently: prompt caching for a stable prefix placed before dynamic content, modular prompts assembled from shared parts, and Skills that package instructions and resources for reuse. The guide's own sample turns on ordering static content first so it can be cached.

prompt cachingstatic prefix before dynamic contentmodular promptsSkills

Practice question for this objective

Free sampleClaude Models, Prompting & Context Engineeringmedium

A government benefits agency assembles each decision-letter prompt from separately owned modules: a tone module, an eligibility module and a statutory-notices module carrying the mandatory appeal-rights paragraph. Each module has its own test suite, and every suite passes. Three weeks ago the communications team added a plain-language module, appended after the others, that limits letters to 200 words. Since then reviewers find the appeal-rights paragraph missing from 12 percent of letters. The model and the original three modules are unchanged. What should the team investigate first?

AWhether a larger model tier would follow each module's instructions more reliably now the assembled prompt is longer
BThe fully assembled prompt for affected letters, checking whether the new word limit conflicts with the mandatory paragraph Correct
CThe statutory-notices module's own test suite, since its passing tests must lack a case for the appeal paragraph
DThe model provider's release notes, since a silent change to the model's weights would explain a sudden regression

Modular prompts can pass module-level tests yet conflict once assembled, so diagnose regressions from the rendered, fully assembled prompt. Modular prompt design lets separate owners maintain parts, but the model only ever sees the assembled whole. A new module that caps length, placed after the others, can lead the model to trade away content another module requires. Because each module is tested in isolation, no suite sees the interaction; inspecting the assembled prompt for the failing cases exposes the conflict and points to a fix such as exempting mandatory text from the limit and adding an assembled-prompt regression test.

Why A is wrong: Tempting because a more capable model can handle more instructions at once. It is wrong because the regression began with a prompt change while the model stayed the same, so a model swap addresses the wrong layer and would leave the conflicting instructions in place.

Why B is correct: Correct. The only change was a new module appended last, and a 200-word ceiling can compete directly with a mandatory paragraph from another module. Reading the rendered prompt as the model receives it is the quickest way to confirm that cross-module conflict.

Why C is wrong: Tempting because missing test coverage is a common cause of silent regressions. It is wrong because that module and its tests are unchanged and passed before and after; a test that exercises one module in isolation cannot reveal a conflict that only exists once modules are combined.

Why D is wrong: Tempting because unexplained regressions are often blamed on the model. It is wrong because the stem states the model is unchanged, and the onset coincides exactly with the new module, which points to the prompt rather than the weights.

See more CCAR-P practice questions, answers explained.

Exam traps in Claude Models, Prompting & Context Engineering

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.

Keep real-time requests but put each product's attributes first and the brand guide after them, so the model reads the item before the rules
Why it is wrong: This is tempting because some teams believe the model attends better to whatever it reads first. It is wrong because placing the variable attributes ahead of the guide means no two requests share a prefix, so nothing can be cached, and real-time requests buy latency the overnight deadline does not need.
Higher traffic causes cached entries to be evicted sooner, so most requests now arrive after the stored prefix has lapsed
Why it is wrong: Tempting because cache lifetime is a real factor in hit rates. It is wrong because rising traffic keeps a cached prefix in frequent use, which works in favour of reuse rather than against it, and nothing about lifetime changed in the release.
The new line conflicts with each product's own instructions, and the model resolves that clash unevenly
Why it is wrong: Conflicts between assembled modules are a common cause of uneven behaviour, so this is a natural hypothesis for a modular prompt. It is wrong because none of the other product sections says anything about invoices, so there is no competing instruction; the model is simply obeying an instruction it now receives.

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