Skip to content

Claude Prompt Engineering Specialist (Self-Directed Track) - Practice Questions

15 questions for the Claude Prompt Engineering Specialist track, weighted toward prompt fundamentals (18%), system prompts and role (16%), reasoning elicitation (16%), then structure, examples, and evaluation.

This is a self-directed study track rather than an Anthropic exam. Prompting guidance shifts between model generations, so verify against current documentation.

Cert page: exams/anthropic/claude-prompt-engineering-specialist/

Question 1

Scenario: What belongs in a prompt, in one principle?

A. What only you know: the audience, the product, environment facts, the quality bar, constraints and their reasons B. Everything that might be relevant, including background the model already knows, so that nothing is left out and no later guesswork is needed C. As few words as possible, because short prompts are cheaper and the model fills in any missing details sensibly and reliably D. Restatements of good behavior, such as being accurate, helpful, and careful, which the model already follows without being told

Answer **Correct: A** **Why:** Context is never cruft; restating trained defaults is. "Be accurate and helpful" adds nothing, while "this goes to compliance reviewers who need the source cited" changes the output. Too-short prompts produce generic results because the model fills gaps with safe defaults.

Question 2

Scenario: A prompt reads "CRITICAL: You MUST ALWAYS verify before answering. NEVER guess. IMPORTANT: be thorough."

A. Over-emphasized: when everything is marked critical the markers carry no information, and an anxious prompt produces a hedging, over-cautious model B. Needs more emphasis, such as capital letters and repeated warnings throughout, so that the model treats every instruction as the highest priority and stops cutting corners C. Correct for all models, because strong wording is a reliable way to get compliance and has no downside on current models D. Well-calibrated emphasis, since each rule carries a marker that tells the model exactly how seriously to take it

Answer **Correct: A** **Why:** Forcefulness was genuinely needed on older, less steerable models. Current models follow the system prompt closely, so the same text over-applies. Emphasis is a tested, scoped fix for one demonstrably underweighted instruction, not a default register.

Question 3

Scenario: A prompt says "try to include a summary if possible" for a summary that is actually required.

A. Repeat the line three times in different places, so that the model sees the requirement often enough to treat it as mandatory B. Hedges are now read literally as permission to skip; state the requirement plainly C. Fine as written, because a polite request is the best way to ask and current models understand that it is actually required D. Add emphasis with capital letters and an exclamation mark, which shouts louder but leaves the hedged wording of the requirement

Answer **Correct: B** **Why:** Over-emphasis and under-emphasis are the same defect from opposite directions, and both come from writing for a model with different steerability. Say exactly what you mean at normal volume.

Question 4

Scenario: A prompt contains "think step by step" and instructions to use <scratchpad> tags.

A. Add more reasoning instructions, such as asking for a longer chain of thought and a final check, to improve accuracy further B. They improve output on every model and should be kept in all prompts, because explicit reasoning cues are always helpful C. Keep them for reliability, since removing reasoning instructions makes the model skip steps and produce lower quality answers, especially on multi-step tasks with many conditions D. On thinking models these are redundant; control reasoning with adaptive thinking and the effort parameter

Answer **Correct: D** **Why:** These are idioms from an era before reasoning was natively trained and configurable. Instructions telling the model *not* to reason are worse still: they can increase leakage of internal markup into the visible response rather than suppressing it.

Question 5

Scenario: A prompt uses XML tags to delimit a document and the instructions.

A. Useful: explicit delimiters make it unambiguous which text is data and which is instruction, especially with long inputs B. Only JSON works for separating data from instructions, so XML tags confuse the model and should be replaced with JSON keys everywhere C. Unnecessary, since the model can always tell where a long document ends D. XML tags are deprecated, and current models ignore them, so plain headings are the only supported way to separate text from instructions

Answer **Correct: A** **Why:** Structural delimiting remains sound practice for separating content from instruction. What has been superseded is using tags to *force output shape*, which structured outputs now handles as a decoding constraint rather than a prompting convention.

Question 6

Scenario: A prompt embeds one gold-standard output example.

A. Risky: examples are the strongest signal in a prompt, so one example freezes its length, tone, and structure; use several deliberately varied examples labeled illustrative B. Add more examples of the same shape, which reinforces the pattern and gives the model a strong template for each output, so every response matches the sample closely in length and tone C. Examples never matter, because the model follows the written instructions alone and ignores any sample outputs in the prompt D. Ideal, because one gold-standard example shows the model precisely what is wanted and avoids ambiguity about the format

Answer **Correct: A** **Why:** The model matches what it sees, so a single example becomes a template rather than an illustration. Examples written for an older model are worse: they freeze that model's behavior into the new one. Keep examples that pin a genuinely format-sensitive output.

Question 7

Scenario: A prompt has fifteen "do not" lines.

A. Add more prohibitions covering every failure seen so far, so that the model has a complete list of behaviors it must avoid B. Describing success beats enumerating failure, and a prohibition can anchor the model toward that failure C. Prohibitions are always wrong, and a prompt should never state what the model must not do D. Comprehensive coverage, since a long list of explicit rules is the safest way to prevent each known failure from recurring, and every incident should add one more line to the list

Answer **Correct: B** **Why:** The test for each line is provenance: does it encode a real constraint with a reason, or is it a workaround for an older model's habit? Keep prohibitions whose failure reproduces on the current model; rewrite the rest as positive statements.

Question 8

Scenario: A system prompt runs to 4,000 words of numbered steps for a judgment task.

A. Convert it to bullets, which makes it easier to read but leaves the same prescriptive structure and the same word count B. More detail is better, so keep adding steps for each case the model gets wrong until every situation is covered C. Over-prescriptive: prompts written for prior models are often too prescriptive for current ones and reduce output quality; state outcomes, constraints, and how to verify, keeping numbered steps only where order truly matters D. Split it into two prompts of 2,000 words each, which divides the length but keeps the over-prescriptive steps in each part

Answer **Correct: C** **Why:** Prompting effort should scale with how far the task is from what the model does naturally. Fragile operations where exactly one sequence is safe keep their exact scripts; open-ended judgment work does better with the goal and the constraints.

Question 9

Scenario: A prompt says "you will be graded on completeness; there are hidden tests."

A. Add the rubric to the prompt as a hidden section, so that the model knows which items will be scored and can plan its answer B. Effective motivation, since telling the model that it is being evaluated makes it try harder and check its work more carefully C. It describes the scoring apparatus instead of the requirement, and pushes effort toward appearing watched; state every requirement the grader checks and never describe the grader D. Remove all quality language from the prompt, including the stated requirements, so that the model is not pushed by any pressure

Answer **Correct: C** **Why:** If a requirement matters, say it as a requirement. Grader vocabulary is a proxy that the model optimizes toward instead of the underlying goal, which is exactly the substitution you do not want.

Question 10

Scenario: How should a prompt change be validated?

A. Ship it and watch for complaints from users, then revise the prompt if people report that the quality has dropped B. Read the new output and judge it, using your impression of a few examples to decide if the change was an improvement C. Ask the model whether the prompt is good, and apply any changes it recommends, since it knows what it responds best to D. An evaluation set with defined metrics, run before and after, ideally changing one thing at a time

Answer **Correct: D** **Why:** Prompting is empirical and intuitions about what helps are unreliable. Asking the model to assess its own prompt is self-report, not measurement. Changing one thing at a time is what lets a regression be attributed to its cause.

Question 11

Scenario: A prompt embeds "current date: {today}" at the top of the system prompt.

A. Dates cannot be provided to the model in a prompt, because it has no clock and any date text would be ignored B. It sits at the front of the cached prefix and invalidates everything after it on every request; inject dynamic context later in the messages instead C. Move it to the end of the system prompt, which still sits inside the cached prefix and changes the cache key each day D. Good practice for temporal grounding, since placing the date first helps the model anchor all later content to the correct day, and the cost of a changing prefix is small compared with the benefit

Answer **Correct: B** **Why:** Caching is a prefix match, and the render order is tools, then system, then messages. Anything varying per request belongs after the last breakpoint. A message at turn five invalidates nothing before turn five.

Question 12

Scenario: A prompt contains lines nobody can justify, added after specific incidents on a retired model.

A. Keep them, they are harmless, because extra instructions never reduce quality and may protect against a failure that returns, so deleting any of them adds risk with no benefit B. Rewrite them more forcefully so that the model gives them clear priority over the rest of the instructions in the prompt C. Treat them as removal candidates: prompts accumulate the union of every generation's workarounds, and specific outdated instructions actively degrade behavior D. Add a comment explaining each, which documents the history for readers but leaves the outdated instructions in the model's context

Answer **Correct: C** **Why:** The question for every emphatic line is which failure, on which model, it prevented, and whether that failure still reproduces. Removal is a hypothesis to test, not a conclusion: probe behavior before and after, and re-add in minimal form if a cut regresses.

Question 13

Scenario: A skill's frontmatter description carries urgent language about when to invoke it.

A. All urgency should be removed from every part of a prompt, including skill descriptions, because strong wording harms any model B. Descriptions should be one word long, since a short name is enough for the model to decide whether to invoke the skill C. Same problem as an over-emphatic system prompt, because urgent wording in any text always causes over-triggering and rigid behavior D. Different: text whose job is routing may legitimately carry calibrated urgency, since skills currently under-trigger; text whose job is behavior should explain rather than shout

Answer **Correct: D** **Why:** Trigger text and behavioral text look identical to a search, so classify by function before editing. Ideally the trigger wording is tuned against a trigger evaluation rather than intuition.

Question 14

Scenario: A response is longer than the product wants.

A. Add "be concise" and hope, which is vague, leaves the model to guess what to cut and how short the response should be, and gives different lengths on different runs B. Give a concrete instruction about what to include and what to drop, and prefer positive examples of the desired concision over telling the model what not to do C. Lower the effort setting, which changes how much the model reasons and does not directly control the length of the final answer D. Cap the output tokens at a small limit, which truncates the answer mid-sentence instead of teaching the model what to include and what to leave out, and so harms quality

Answer **Correct: B** **Why:** Numeric output caps tuned against an older model's verbosity starve reasoning on hard problems, and truncate rather than shorten. Lowering effort changes reasoning depth without reliably changing visible length. Selectivity is the lever: drop what would not change what the reader does next.

Question 15

Scenario: A prompt is being written for a new model generation.

A. Reuse the old prompt unchanged, because a prompt that worked well on the previous model will also work on its successor, and rewriting it only adds the risk of new errors B. Re-baseline from the old prompt: remove workarounds the new model no longer needs, and add guidance for its new behaviors, since prompts are per-model artifacts C. Always shorten it, because newer models need fewer instructions and any extra text will lower the quality of the output, so every migration should cut the prompt by half or more D. Always lengthen it, because newer models can use more detail, and adding extra guidance can only improve what they produce

Answer **Correct: B** **Why:** Matching a prompt to a new model sometimes means adding text, not only cutting it. A line that is load-bearing on one generation is cruft on the next, and the reverse, which is why each model release is a trigger to re-audit rather than a reason to reuse.

Where to go deeper