Prompt Engineering & Structured Output

Question 1 · 4.1

A retailer’s CI pipeline runs Claude to review every pull request. Engineers report that most comments are about variable naming, missing docstrings and possible null values that the type checker already rules out, and they have started ignoring the bot.

Which change most directly improves precision without losing real bug reports?

• AAdd “Only report issues you are highly confident about.”
• BSwitch to a larger model so its judgement of what matters improves.
• CDefine bug categories to report, a skip list, and a file:line evidence rule.
• DPrefix the instructions with “CRITICAL: you MUST NOT produce false positives.”

Question 2 · 4.1

A hospital triage assistant labels incoming patient-portal messages as urgent or routine. Staff find that messages with words like “severe” or “emergency” are always marked urgent, even when they describe an old, resolved problem.

What should the prompt add?

• AA definition of urgent based on current impact, not alarming words.
• BAn instruction to lower the urgency of any message containing the word “emergency”.
• CA request that the model output a confidence score from 1 to 10.
• DA reminder that false positives waste staff time, written in capitals.

Question 3 · 4.1

An insurer uses Claude to flag claims that may need a fraud review. The team wants fewer false alarms but cannot afford to miss genuine cases.

Which two design choices fit that goal? (Select 2.)

• ATell the model to flag a claim only when it is certain of fraud.
• BList the specific indicators that justify a flag and require the finding to cite the claim fields that show them.
• CAsk for every candidate with a defined severity and category, then filter in code or a second pass.
• DRemove examples of borderline claims from the prompt to avoid biasing the model.
• EGive the model a longer list of fraud-related keywords to watch for.

Question 4 · 4.2

An insurer’s claim classifier has a detailed system prompt defining each claim type. Reviewers find it consistently puts storm-damaged vehicles under “weather” one day and “motor” the next, depending on wording.

What is the most effective next step?

• ARewrite the definitions of “weather” and “motor” at greater length.
• BAdd ten more examples of clear-cut motor claims.
• CTell the model to think very carefully about borderline claims.
• DAdd examples of storm-damaged vehicles, labelled and with a one-line reason.

Question 5 · 4.2

An extraction prompt pulls invoice fields from supplier PDFs. All four examples in the prompt show complete invoices. On invoices with no purchase-order number, the model sometimes returns a plausible-looking but invented PO number.

Which change addresses the cause?

• AAdd an example invoice with no PO, extracted as po_number: null.
• BRemove all the examples so the model is no longer biased by them.
• CMake po_number a required string in the schema.
• DInstruct the model in capitals never to invent data.

Question 6 · 4.2

An internal IT agent calls a create_ticket tool. Calls are valid JSON, but dates arrive in three different formats and the optional reporter object is filled in with guessed values.

Which two changes are most appropriate? (Select 2.)

• AAdd input_examples showing a full call and a minimal call with realistic values.
• BState the expected date format in the due_date property description.
• CAdd placeholder examples such as {"due_date": "string"}.
• DInclude examples that omit required fields to show flexibility.
• EForce tool_choice to create_ticket on every request.

Question 7 · 4.3

A claims pipeline asks Claude to “respond only with valid JSON” describing each claim. Roughly one response in fifty has a sentence before the JSON or a missing key, and the loader crashes.

What is the most reliable fix?

• ARepeat the JSON instruction at the end of the prompt in capitals.
• BStrip everything before the first brace with a regular expression.
• CDefine the claim as a tool’s input_schema with strict mode and force that tool.
• DAdd three examples of well-formed claim JSON to the prompt.

Question 8 · 4.3

A mailroom agent receives invoices, receipts and purchase orders mixed together, unlabelled. There is a separate extraction tool for each type, and every document must produce structured data.

Which tool_choice setting fits best?

• Aauto, so Claude can decide whether a tool is needed.
• Bany, so Claude must call one of the tools and chooses which.
• Ctool naming the invoice extractor, then re-run if it fails.
• Dnone, then parse the text response into the right schema.

Question 9 · 4.3

After moving to strict tool use, an extraction pipeline never produces malformed JSON. An audit finds that on contracts with no renewal date, renewal_date is filled with the signature date. The field is a required string.

What change addresses the root cause?

• AMake renewal_date nullable and show in an example that null means “not stated”.
• BTurn off strict mode so the model can omit the field.
• CAdd a pattern constraint so only valid dates are accepted.
• DAsk the model to double-check dates before answering.

Question 10 · 4.3

A strict JSON output response comes back with stop_reason of max_tokens. What should the application assume?

• AThe JSON is valid, because strict mode guarantees it.
• BThe model refused and should not be retried.
• CThe payload may be incomplete; retry with a higher max_tokens.
• DThe schema is too complex and must be simplified.

Question 11 · 4.4

A shipping company’s extraction pipeline validates each bill of lading. When the container count does not match the listed containers, it resends the same request, up to three times. Most failing documents fail all three attempts.

What change will most improve the retry success rate?

• ARaise the retry limit from three attempts to ten.
• BSend the failed output and the exact mismatch with the retry.
• CSwitch the retry to a different, larger model.
• DRemove the container-count check to stop the failures.

Question 12 · 4.4

An HR system extracts employee start dates from signed offer letters. Some older letters were signed without a stated start date. The validator rejects records without a date and triggers a retry.

How should the pipeline treat these letters?

• AKeep retrying until a date is produced, up to a limit.
• BFill in the signature date as a default start date.
• CTell the model to infer the most likely start date.
• DAllow null for a missing start date and route it to HR to confirm.

Question 13 · 4.4

A bank extracts loan applications with a strict extraction tool, so every response matches the schema. Auditors find cases where monthly income exceeds annual income, and where the applicant’s name was taken from the referee section.

Which two measures address these errors? (Select 2.)

• AAdd code checks for cross-field rules such as annual ≥ 12 × monthly income.
• BAdd a source_text field per key value and verify it appears in the relevant section.
• CTurn strict mode off so the model has more freedom.
• DAsk the model to confirm that it is confident in each value.
• ERetry every application twice and keep the most common answer.

Question 14 · 4.4

An Agent SDK run with output_format set ends with subtype error_max_structured_output_retries. What does this mean?

• AThe agent used up its turn limit while calling tools.
• BNo schema-valid output was produced within the retry limit.
• CThe schema itself was invalid, so it was ignored.
• DThe output is valid but exceeded max_tokens.

Question 15 · 4.5

A team runs two Claude workloads: a check that must pass before a pull request can merge, and an overnight job that summarises every support ticket from the previous day for a morning dashboard.

Which assignment of APIs is most appropriate?

• ABoth on the Batches API, to halve cost across the board.
• BBoth synchronous, because batches cannot use tools.
• CMerge check synchronous; overnight summaries on the Batches API.
• DMerge check on batches with polling every minute; summaries synchronous.

Question 16 · 4.5

A batch of 20,000 extraction requests ends with 19,640 succeeded, 300 errored with invalid_request_error, and 60 expired. The engineer’s first instinct is to resubmit the entire batch.

What is the best next step?

• AResubmit all 20,000 requests so the results are consistent.
• BRetry the 360 failed requests unchanged in a new batch.
• CDiscard the failed items, since 98% coverage is enough.
• DFix the 300 invalid requests, then resubmit them with the 60 expired ones.

Question 17 · 4.5

A compliance team must have each uploaded contract reviewed within 30 hours of upload. They want to use the Batches API to cut cost, and documents arrive at random times throughout the day.

How often must they submit batches to guarantee the deadline?

• AAt least every 6 hours, because a batch can take up to 24 hours.
• BOnce a day, because most batches finish within an hour.
• COnce every 30 hours, matching the deadline.
• DIt cannot be done; batches are unsuitable for any deadline.

Question 18 · 4.5

Why should each request in a batch carry a meaningful custom_id?

• AIt sets the priority with which requests are processed.
• BResults may return in any order, so it is how results are matched.
• CIt enables prompt caching across requests in the batch.
• DIt is used as the cache key for the compiled schema grammar.

Question 19 · 4.6

An engineering team’s code-generation agent writes a module, then in the same conversation is asked “Review the code you just wrote for bugs.” It almost always reports no significant issues, yet human reviewers keep finding logic errors.

What change is most likely to improve the review?

• ATell the agent to review “extremely carefully” before answering.
• BGive the review to a separate instance that sees only the code and criteria.
• CRaise max_tokens so the self-review can be more detailed.
• DAsk the agent to rate its confidence in the code from 1 to 10.

Question 20 · 4.6

A single-pass review of a 35-file pull request returns a handful of style comments and misses that a shared function’s return type changed while three callers were not updated.

Which architecture best addresses this?

• AUse a model with a larger context window for the single pass.
• BReview each file independently and merge the comments.
• CRun the same single-pass review three times and combine the results.
• DPer-file passes, then an integration pass over interfaces and call sites.

Question 21 · 4.6

A legal-tech company’s contract reviewer produces too many false positives, but lawyers insist it must not miss genuinely risky clauses.

Which two design changes fit these goals? (Select 2.)

• AKeep the finding pass broad, then add a verification pass that tries to refute each finding.
• BTell the finder to report only issues it is certain about.
• CGive the verifier explicit criteria for what counts as a risky clause and require quoted evidence.
• DHave the finder re-read its own findings in the same conversation.
• EMerge finding and verification into one longer prompt to save calls.

Question 22 · 4.6

Three independent instances extract the dosage of a high-risk drug from the same record. Two say 5 mg, one says 50 mg. What should the pipeline do?

• AAccept 5 mg, because two of the three instances agree.
• BSend it to a human reviewer with all three readings.
• CRerun until all three instances agree.
• DDiscard the record because extraction is unreliable.