Design an enterprise registry for 600 production prompt templates owned by 14 teams, with 99.9% availability and a 10-minute publication SLA.

I would make the registry the typed source of truth for prompt identity, ownership, versions, approvals, and evaluation evidence. - Each immutable version stores its template, variable schema, output contract, model variants, owner, risk tier, and links to the exact eval report. - Draft, reviewed, approved, canary, active, and deprecated are explicit states; only signed approved versions can enter a production bundle. - The 10-minute SLA covers validation and publication, while cached active bundles keep products running during registry downtime.

Define an immutable prompt bundle manifest for 180 templates, three providers, six languages, and a rollback target below five minutes.

I would release one content-addressed manifest that resolves every prompt dependency before traffic reaches it. - The manifest pins template hashes, locale variants, variable and output schemas, policy fragments, examples, model adapter revisions, and evaluation dataset versions. - Build validation rejects missing language files, mutable aliases, incompatible schemas, and unresolved inheritance across all 180 templates. - Production traces carry the manifest ID, and rollback changes one active pointer to the prior signed bundle within five minutes.

Design template composition and inheritance for 400 prompts across 12 products, while limiting any rendered prompt to 6,000 input tokens and two inheritance levels.

I would use explicit composition with a shallow directed dependency graph, not unrestricted textual inheritance. - Shared policy, brand, task, and locale fragments have typed interfaces; product prompts select them and may override only declared extension points. - The compiler detects cycles, shadowed variables, conflicting instructions, duplicate fragments, and the two-level depth limit before publication. - A rendered snapshot and token count are evaluated per product and model, so a shared edit cannot silently push a prompt above 6,000 tokens.

Create role and provider adapters for 90 prompts that must behave consistently on GPT, Claude, and Gemini, with less than a two-point quality gap and a $0.03 request budget.

I would keep task intent provider-neutral and compile it through provider-specific role and feature adapters. - The common contract defines instruction priority, variables, examples, tools, and output schema, while adapters map system roles, tool syntax, and structured-output features honestly. - Unsupported capabilities fail compatibility checks instead of being approximated silently, and each provider variant has paired evaluation against the same cases. - I would approve only variants within the two-point quality bound and $0.03 budget, documenting intentional behavioral differences rather than forcing identical text.

Plan a 30-day migration of 250 customer-support prompts from GPT to Claude and Gemini across English, Spanish, and Japanese, with no more than 1% task-quality regression.

I would use a hierarchical evaluation design sized to detect the 1% bound, not spread 6,000 cases evenly across 250 prompts. - A compatibility matrix covers role hierarchy, refusals, tool descriptions, JSON constraints, context ordering, stop behavior, and language-specific instruction following. - Power analysis sets an adequate paired sample for aggregate risk and critical provider, language, intent, and prompt-family slices, with oversampling for rare high-risk tasks. - Shadow traffic and 5%, 20%, and 50% stages precede full migration, with the old prompt-provider bundle ready for rollback throughout 30 days.

Design a prompt compiler and linter for 1,000 templates changed by 20 teams, with CI feedback below 90 seconds and fewer than 0.5% false blocks.

I would compile templates into deterministic rendered artifacts and split fast static checks from targeted behavioral tests. - Static rules catch undefined variables, dead branches, role-order violations, duplicate instructions, forbidden secrets, schema mismatches, and token-limit breaches. - Changed-fragment analysis selects affected prompts and locales, keeping the common CI path below 90 seconds while a broader suite runs before promotion. - Every lint rule has positive and negative fixtures, severity, owner, and suppression expiry so false blocks stay below 0.5% without permanent bypasses.

Specify variable typing and escaping for 300 templates receiving user text up to 20,000 characters in six languages, with zero instruction-boundary escapes in 10,000 adversarial tests.

I would give every variable a declared type, trust class, size bound, and canonical serializer instead of interpolating raw strings. - Types distinguish plain text, trusted policy, quoted evidence, locale, enum, JSON, and tool arguments; the compiler rejects unsafe concatenation and structural escapes. - Safe rendering prevents data from changing template structure, but it does not prove behavioral resistance when the model reads malicious instructions inside valid data, so a separate injection suite tests that behavior. - Server code independently authorizes every action, constrains tool and argument allowlists, and treats model output as an untrusted request before execution.

Design prompt and output contract schemas for a 24-field claims workflow processing 2 million documents per month with 99.95% contract validity.

I would version the input, prompt, and output contracts together and keep semantic invariants explicit outside prose. - The input schema defines required evidence, nullable fields, units, locale, and policy revision; the prompt may reference only declared fields. - The output schema pins types, enums, evidence handles, confidence semantics, and refusal reasons, while deterministic validation checks cross-field rules. - Contract tests cover upgrades and downgrades, and release requires 99.95% validity plus field-level accuracy on production-shaped claims.

Architect a structured-output prompt fleet of 220 extractors serving 8 million calls per day, with 99.99% valid JSON and at most one repair attempt.

I would standardize a schema-first prompt pattern and use each provider's strongest constrained-output mechanism. - Shared instructions define evidence use, null handling, forbidden inference, and field descriptions, while each extractor supplies a versioned JSON Schema. - Syntax, schema, and semantic failures are separate metrics; only repairable formatting failures receive one retry with the exact validation error. - Fleet gates require 99.99% JSON validity, field accuracy floors, bounded output tokens, and no hidden retry increase across all active model variants.

Design a five-stage prompt chain for 500,000 contract reviews per month, with typed stage contracts, p95 below 12 seconds, and a $0.18 review budget.

I would make each stage produce the minimum typed artifact needed by the next stage and reject invalid transitions early. - Classification, clause extraction, policy comparison, risk synthesis, and final explanation each have independent schemas, prompts, token caps, and eval sets. - Stage outputs carry evidence handles and confidence, while prose summaries never replace exact extracted clauses or policy identifiers. - Parallelizable stages run together, but the release gate measures end-to-end accuracy, stage failure attribution, p95 below 12 seconds, and cost below $0.18.

Separate planner and executor prompts for an agent limited to eight tool calls, 45 seconds, 12,000 total prompt tokens, and two replans.

I would give planning and execution different contracts, context, and authority, with budgets stated in both prompts and enforced externally. - The planner returns goals, dependencies, required evidence, stop conditions, and a bounded plan without tool results or permission to execute. - The executor receives one approved step, typed tool descriptions, prior observations, and a remaining-budget summary, but cannot rewrite the overall objective. - Eval cases score plan sufficiency, invalid actions, needless replans, completion within eight calls and 45 seconds, and total prompt use below 12,000 tokens.

Write the architecture for ReAct policy prompts used by 40 workflows, with a maximum of six reasoning-action cycles and fewer than 0.2% policy-violating action proposals.

I would put the ReAct loop under a runtime state machine and use the prompt only to guide proposals within that control boundary. - The state machine counts every reasoning-action cycle, accepts explicit completion, refusal, or escalation states, and hard-stops after the sixth cycle. - Every proposed action must pass schema validation and an external policy authorization check before execution; prompt wording never grants authority. - Typed untrusted observations and adversarial tests cover repeated actions, unsupported conclusions, and policy-violating proposals, with 0.2% as the release ceiling.

Govern 75 tool descriptions shared by 10 agent teams, with a 4,000-token tool-description budget and 98% correct tool selection.

I would treat tool descriptions as versioned prompt interfaces owned with the underlying operations. - Each description states one purpose, eligibility, required arguments, side effects, exclusions, and two discriminating examples, avoiding overlapping marketing language. - A linter detects duplicate intent, contradictory argument guidance, missing side-effect labels, and total rendered size above 4,000 tokens. - A shared confusion set measures selection and abstention across all 75 tools, and any schema or description change must keep correct selection above 98%.

Design answer, citation, and conflict prompts for a policy assistant handling 3 million questions per month, with 97% citation precision and support for documents from three authority levels.

I would make authority and evidence handling explicit in separate prompt sections and require claim-level citation handles in the output contract. - The prompt tells the model to answer only from supplied spans, cite every factual claim, and abstain when no span supports it. - Authority metadata defines which of the three levels wins; unresolved conflicts must be presented with both citations rather than silently reconciled. - Evaluation separates answer correctness, citation precision, citation coverage, conflict handling, and refusal accuracy, with 97% citation precision as a release floor.

Order a 48,000-token prompt containing 30 evidence sections for a legal summarizer, with at least 90% fact retention and output below 2,000 tokens.

I would protect instructions and output rules, then organize evidence to reduce lost-in-the-middle failures rather than preserve ingestion order. - Task, definitions, and output schema come first; evidence gets stable IDs, headings, dates, and authority labels, with the most decision-critical spans near prompt boundaries. - A short evidence map precedes the full sections, while duplicated boilerplate is removed and no model-written summary replaces quoted legal text. - Position-swap tests across all 30 sections measure fact retention, citation support, and the 2,000-token output cap before the layout is approved.

Design a constitutional prompt hierarchy for 120 templates across four risk tiers, with policy updates published within 24 hours and zero unauthorized lower-level overrides.

I would define a fixed precedence chain from enterprise constitution to domain policy, task instructions, tenant preferences, and user input. - Each layer has allowed decisions and protected clauses; lower layers can specialize declared fields but cannot replace safety, legal, or data-use rules. - The compiler renders the effective hierarchy, flags contradictions, and proves that every one of the 120 templates includes the current policy revision. - Policy fixtures test direct, indirect, multilingual, and conflicting instructions, while signed bundle publication meets the 24-hour update target.

Build a safety prompt library for 16 product teams and eight harm categories, with critical-case recall above 99% in English and Spanish.

I would provide small risk-specific prompt modules with shared refusal and safe-completion contracts, not one giant universal safety paragraph. - Each of the eight categories defines prohibited assistance, allowed high-level help, escalation language, and counterexamples reviewed by policy specialists. - Products compose only modules matching their capabilities and risk tier, with locale variants reviewed independently rather than translated mechanically. - A sealed adversarial set gates English and Spanish critical recall above 99%, while over-refusal is measured on legitimate neighboring requests.

Design an automated prompt-injection and red-team suite for 200 prompts, 25 attack families, six languages, and a four-hour nightly budget.

I would combine fixed regression attacks with generated mutations and score concrete behavior failures rather than keyword detection. - The 25 families cover direct, indirect, encoded, role-confusion, delimiter, multilingual, tool-description, and long-context attacks with expected safe outcomes. - Risk-based sampling runs all critical prompts nightly and rotates lower-risk combinations so the complete 200-prompt matrix finishes over one week. - Nightly execution stays under four hours through changed-prompt selection, and any unauthorized disclosure or action proposal blocks promotion regardless of average score.

Define tenant customization boundaries for 2,500 customers editing 140 base prompts, with 99.95% safe rendering and a 3,000-token customization cap.

I would accept tenant customization only through an allowlisted typed schema rendered into fixed slots that are always treated as untrusted data. - The schema permits bounded tone, terminology, examples, and approved business-rule fields; unknown fields, structural text, and content above 3,000 tokens are rejected. - Safety policy, tool authority, role hierarchy, variable boundaries, and output contracts are assembled outside tenant text, so customization cannot occupy or redefine those positions. - Linting for conflicts, secrets, and injection patterns is supplemental, followed by tenant and global safety evaluations and fallback to the last approved version.

Operate a multilingual fleet of 160 prompts in English, Russian, French, German, Portuguese, and Spanish, with language quality within three points and releases inside 48 hours.

I would treat each locale as a behavior variant with shared intent and independent linguistic evaluation. - Canonical requirements, variables, examples, and output contracts are language-neutral, while native prompt text may adapt formality, syntax, and culturally ambiguous terms. - Translation memory helps consistency, but native reviewers own high-risk prompts and a parity check blocks missing or stale locale variants. - A 48-hour release uses automated six-language regression plus sampled native review, and no locale may trail the approved baseline by more than three points.

Prompt Engineer interview questions

100 real questions with model answers and explanations for Staff Prompt Engineer candidates.

See a Prompt Engineer resume example →

Practice with flashcards

Spaced repetition · Hunter Pass

Questions

promptingregistriestemplates

I would make the registry the typed source of truth for prompt identity, ownership, versions, approvals, and evaluation evidence.

Each immutable version stores its template, variable schema, output contract, model variants, owner, risk tier, and links to the exact eval report.
Draft, reviewed, approved, canary, active, and deprecated are explicit states; only signed approved versions can enter a production bundle.
The 10-minute SLA covers validation and publication, while cached active bundles keep products running during registry downtime.

Why interviewers ask this: The interviewer is checking whether the candidate can turn hundreds of prompts and many teams into an auditable lifecycle with a resilient serving boundary.

promptingrollbackprompt-bundle

I would release one content-addressed manifest that resolves every prompt dependency before traffic reaches it.

The manifest pins template hashes, locale variants, variable and output schemas, policy fragments, examples, model adapter revisions, and evaluation dataset versions.
Build validation rejects missing language files, mutable aliases, incompatible schemas, and unresolved inheritance across all 180 templates.
Production traces carry the manifest ID, and rollback changes one active pointer to the prior signed bundle within five minutes.

Why interviewers ask this: A strong answer makes prompt behavior reproducible as a complete bundle rather than a collection of independently changing strings.

promptingtokensoop

I would use explicit composition with a shallow directed dependency graph, not unrestricted textual inheritance.

Shared policy, brand, task, and locale fragments have typed interfaces; product prompts select them and may override only declared extension points.
The compiler detects cycles, shadowed variables, conflicting instructions, duplicate fragments, and the two-level depth limit before publication.
A rendered snapshot and token count are evaluated per product and model, so a shared edit cannot silently push a prompt above 6,000 tokens.

Why interviewers ask this: The interviewer is evaluating whether reuse remains understandable, bounded, and testable at fleet scale.

prompting

I would keep task intent provider-neutral and compile it through provider-specific role and feature adapters.

The common contract defines instruction priority, variables, examples, tools, and output schema, while adapters map system roles, tool syntax, and structured-output features honestly.
Unsupported capabilities fail compatibility checks instead of being approximated silently, and each provider variant has paired evaluation against the same cases.
I would approve only variants within the two-point quality bound and $0.03 budget, documenting intentional behavioral differences rather than forcing identical text.

Why interviewers ask this: A strong answer separates stable prompt intent from provider semantics without pretending the providers are interchangeable.

promptingmigrations

I would use a hierarchical evaluation design sized to detect the 1% bound, not spread 6,000 cases evenly across 250 prompts.

A compatibility matrix covers role hierarchy, refusals, tool descriptions, JSON constraints, context ordering, stop behavior, and language-specific instruction following.
Power analysis sets an adequate paired sample for aggregate risk and critical provider, language, intent, and prompt-family slices, with oversampling for rare high-risk tasks.
Shadow traffic and 5%, 20%, and 50% stages precede full migration, with the old prompt-provider bundle ready for rollback throughout 30 days.

Why interviewers ask this: The interviewer is checking whether cross-provider migration accounts for prompt semantics and uses statistically adequate hierarchical coverage rather than thin per-prompt averages.

promptinglintingcompiler

I would compile templates into deterministic rendered artifacts and split fast static checks from targeted behavioral tests.

Static rules catch undefined variables, dead branches, role-order violations, duplicate instructions, forbidden secrets, schema mismatches, and token-limit breaches.
Changed-fragment analysis selects affected prompts and locales, keeping the common CI path below 90 seconds while a broader suite runs before promotion.
Every lint rule has positive and negative fixtures, severity, owner, and suppression expiry so false blocks stay below 0.5% without permanent bypasses.

Why interviewers ask this: A strong answer treats prompt quality tooling as a compiler with deterministic diagnostics and measured developer usability.

testing

I would give every variable a declared type, trust class, size bound, and canonical serializer instead of interpolating raw strings.

Types distinguish plain text, trusted policy, quoted evidence, locale, enum, JSON, and tool arguments; the compiler rejects unsafe concatenation and structural escapes.
Safe rendering prevents data from changing template structure, but it does not prove behavioral resistance when the model reads malicious instructions inside valid data, so a separate injection suite tests that behavior.
Server code independently authorizes every action, constrains tool and argument allowlists, and treats model output as an untrusted request before execution.

Why interviewers ask this: The interviewer is evaluating whether the candidate separates structural rendering safety from behavioral prompt injection resistance and keeps authorization outside the model.

promptingschemadesign

I would version the input, prompt, and output contracts together and keep semantic invariants explicit outside prose.

The input schema defines required evidence, nullable fields, units, locale, and policy revision; the prompt may reference only declared fields.
The output schema pins types, enums, evidence handles, confidence semantics, and refusal reasons, while deterministic validation checks cross-field rules.
Contract tests cover upgrades and downgrades, and release requires 99.95% validity plus field-level accuracy on production-shaped claims.

Why interviewers ask this: A strong answer connects prompt wording to typed input and output obligations that can evolve safely.

promptingjson-outputstructured-output

I would standardize a schema-first prompt pattern and use each provider's strongest constrained-output mechanism.

Shared instructions define evidence use, null handling, forbidden inference, and field descriptions, while each extractor supplies a versioned JSON Schema.
Syntax, schema, and semantic failures are separate metrics; only repairable formatting failures receive one retry with the exact validation error.
Fleet gates require 99.99% JSON validity, field accuracy floors, bounded output tokens, and no hidden retry increase across all active model variants.

Why interviewers ask this: The interviewer is checking whether structured-output reliability is governed as a prompt family rather than repaired ad hoc per template.

promptingdesign

I would make each stage produce the minimum typed artifact needed by the next stage and reject invalid transitions early.

Classification, clause extraction, policy comparison, risk synthesis, and final explanation each have independent schemas, prompts, token caps, and eval sets.
Stage outputs carry evidence handles and confidence, while prose summaries never replace exact extracted clauses or policy identifiers.
Parallelizable stages run together, but the release gate measures end-to-end accuracy, stage failure attribution, p95 below 12 seconds, and cost below $0.18.

Why interviewers ask this: A strong answer prevents a prompt chain from becoming an opaque sequence of free-form model calls.

promptingtokensagents

I would give planning and execution different contracts, context, and authority, with budgets stated in both prompts and enforced externally.

The planner returns goals, dependencies, required evidence, stop conditions, and a bounded plan without tool results or permission to execute.
The executor receives one approved step, typed tool descriptions, prior observations, and a remaining-budget summary, but cannot rewrite the overall objective.
Eval cases score plan sufficiency, invalid actions, needless replans, completion within eight calls and 45 seconds, and total prompt use below 12,000 tokens.

Why interviewers ask this: The interviewer is evaluating whether prompt separation reduces authority confusion and makes agent budgets measurable.

promptingarchitecturereact

I would put the ReAct loop under a runtime state machine and use the prompt only to guide proposals within that control boundary.

The state machine counts every reasoning-action cycle, accepts explicit completion, refusal, or escalation states, and hard-stops after the sixth cycle.
Every proposed action must pass schema validation and an external policy authorization check before execution; prompt wording never grants authority.
Typed untrusted observations and adversarial tests cover repeated actions, unsupported conclusions, and policy-violating proposals, with 0.2% as the release ceiling.

Why interviewers ask this: A strong answer separates prompt guidance from runtime enforcement through a counted state machine, action validation, and external authorization.

tokensagentstool-descriptions

I would treat tool descriptions as versioned prompt interfaces owned with the underlying operations.

Each description states one purpose, eligibility, required arguments, side effects, exclusions, and two discriminating examples, avoiding overlapping marketing language.
A linter detects duplicate intent, contradictory argument guidance, missing side-effect labels, and total rendered size above 4,000 tokens.
A shared confusion set measures selection and abstention across all 75 tools, and any schema or description change must keep correct selection above 98%.

Why interviewers ask this: The interviewer is checking whether tool descriptions are managed as behavior-critical prompt assets with fleet-level ambiguity tests.

citationspromptingdesign

I would make authority and evidence handling explicit in separate prompt sections and require claim-level citation handles in the output contract.

The prompt tells the model to answer only from supplied spans, cite every factual claim, and abstain when no span supports it.
Authority metadata defines which of the three levels wins; unresolved conflicts must be presented with both citations rather than silently reconciled.
Evaluation separates answer correctness, citation precision, citation coverage, conflict handling, and refusal accuracy, with 97% citation precision as a release floor.

Why interviewers ask this: A strong answer focuses on grounded answer behavior and source precedence without drifting into retrieval-system design.

promptingtokensretention

I would protect instructions and output rules, then organize evidence to reduce lost-in-the-middle failures rather than preserve ingestion order.

Task, definitions, and output schema come first; evidence gets stable IDs, headings, dates, and authority labels, with the most decision-critical spans near prompt boundaries.
A short evidence map precedes the full sections, while duplicated boilerplate is removed and no model-written summary replaces quoted legal text.
Position-swap tests across all 30 sections measure fact retention, citation support, and the 2,000-token output cap before the layout is approved.

Why interviewers ask this: The interviewer is evaluating whether long-context prompt ordering is designed and tested as an information-placement problem.

promptingdesign

I would define a fixed precedence chain from enterprise constitution to domain policy, task instructions, tenant preferences, and user input.

Each layer has allowed decisions and protected clauses; lower layers can specialize declared fields but cannot replace safety, legal, or data-use rules.
The compiler renders the effective hierarchy, flags contradictions, and proves that every one of the 120 templates includes the current policy revision.
Policy fixtures test direct, indirect, multilingual, and conflicting instructions, while signed bundle publication meets the 24-hour update target.

Why interviewers ask this: A strong answer turns constitutional prompting into explicit precedence and override rules that can be compiled and tested.

prompting

I would provide small risk-specific prompt modules with shared refusal and safe-completion contracts, not one giant universal safety paragraph.

Each of the eight categories defines prohibited assistance, allowed high-level help, escalation language, and counterexamples reviewed by policy specialists.
Products compose only modules matching their capabilities and risk tier, with locale variants reviewed independently rather than translated mechanically.
A sealed adversarial set gates English and Spanish critical recall above 99%, while over-refusal is measured on legitimate neighboring requests.

Why interviewers ask this: The interviewer is checking whether safety prompts are modular, localized, and evaluated for both misses and unnecessary refusals.

promptinginjectionred-teaming

I would combine fixed regression attacks with generated mutations and score concrete behavior failures rather than keyword detection.

The 25 families cover direct, indirect, encoded, role-confusion, delimiter, multilingual, tool-description, and long-context attacks with expected safe outcomes.
Risk-based sampling runs all critical prompts nightly and rotates lower-risk combinations so the complete 200-prompt matrix finishes over one week.
Nightly execution stays under four hours through changed-prompt selection, and any unauthorized disclosure or action proposal blocks promotion regardless of average score.

Why interviewers ask this: A strong answer creates a scalable adversarial evaluation system centered on prompt behavior and explicit failure outcomes.

promptingtokens

I would accept tenant customization only through an allowlisted typed schema rendered into fixed slots that are always treated as untrusted data.

The schema permits bounded tone, terminology, examples, and approved business-rule fields; unknown fields, structural text, and content above 3,000 tokens are rejected.
Safety policy, tool authority, role hierarchy, variable boundaries, and output contracts are assembled outside tenant text, so customization cannot occupy or redefine those positions.
Linting for conflicts, secrets, and injection patterns is supplemental, followed by tenant and global safety evaluations and fallback to the last approved version.

Why interviewers ask this: The interviewer is evaluating whether customization is structurally confined by typed slots while policy and authority remain outside untrusted tenant text.

prompting

I would treat each locale as a behavior variant with shared intent and independent linguistic evaluation.

Canonical requirements, variables, examples, and output contracts are language-neutral, while native prompt text may adapt formality, syntax, and culturally ambiguous terms.
Translation memory helps consistency, but native reviewers own high-risk prompts and a parity check blocks missing or stale locale variants.
A 48-hour release uses automated six-language regression plus sampled native review, and no locale may trail the approved baseline by more than three points.

Why interviewers ask this: A strong answer balances shared prompt semantics with genuine locale-specific authorship and release evidence.

Locked questions

21
Design prompt controls for a regulated lending assistant processing 400,000 applications per month, with 100% policy citation coverage and a seven-year audit requirement.
citationspromptingdesign
22
Minimize PII in prompts for 12 million support requests per month, with fewer than 0.01% unnecessary PII fields and support for four regions.
promptingpii
23
Cut a 9,500-token support prompt to 5,000 tokens for 20 million monthly calls while losing less than one quality point and saving at least $120,000 per month.
promptingtokens
24
Lay out a caching-friendly prompt for 70% of 8 million daily calls sharing a 2,800-token prefix, without reducing quality by more than 0.5 points.
promptingtokenscaching
25
Manage per-model variants for 320 prompts across five model families, with fewer than 10% duplicated lines and a three-day model onboarding SLA.
promptingonboarding
26
Design fallback prompt compatibility for 85 workflows that may switch models within 30 seconds, with 99.5% valid outputs and no unsupported tool instructions.
promptingdesign
27
Architect a prompt evaluation platform for 900 templates, 15 teams, 50,000 labeled cases, and a full regression deadline of six hours.
promptingestimationllm-eval
28
Build a 10,000-case golden set from 40 million monthly interactions across six languages and four risk tiers, with quarterly refresh limited to 1,500 labels.
29
Design rubric architecture for 60 prompt tasks scored by 80 reviewers, with at least 0.75 inter-rater agreement and no rubric longer than eight criteria.
promptingarchitecturedesign
30
Calibrate judge prompts for 25 tasks and four languages when judge-human agreement is 84% overall but 62% in Japanese, with a two-week deadline.
promptingestimationllm-judge
31
Design human adjudication for 6,000 monthly prompt-eval disagreements, using 12 domain experts with a 48-hour SLA and a $35,000 monthly budget.
promptingdesignconflict
32
Build an A/B prompt experiment platform for 30 million monthly sessions, 25 concurrent experiments, and a 1% maximum safety-regression guardrail.
promptingsessionsllm-safety
33
Set quality gates for 500 prompt releases per month across three risk tiers, with CI under 15 minutes and critical regressions below 0.1%.
promptingquality-gates
34
Design prompt observability and lineage for 100 million calls per month, with 30-day diagnostics, 13-month aggregate retention, and four regional data boundaries.
promptingdesignlineage
35
Detect semantic behavior drift across 300 stable prompts when traffic changes 8% weekly, with alerts within 24 hours and fewer than five false alarms per month.
promptingalertingiac
36
Plan staged rollout and rollback for a prompt bundle touching 220 templates and 18 teams, with 99.95% availability and rollback below three minutes.
promptingrollbackprompt-bundle
37
Drive prompt standards across 22 teams and 700 templates in one quarter, with 80% adoption and no more than two hours of training per engineer.
promptingdecision-making
38
Design prompt change review for 450 monthly pull requests across 16 teams, with a four-hour median review time and two mandatory reviewers for high-risk prompts.
promptingdesigncode-review
39
Define ownership and deprecation for 900 prompts when 18% have no active owner, with a six-month cleanup target and zero unannounced removals.
promptingownership
40
Create an emergency prompt hotfix process for 24-hour operations, with mitigation in 15 minutes, full review in four hours, and 20 product teams.
promptingconcurrency
41
Govern code-generation prompts used by 1,200 developers in four languages, with 85% sandbox pass rate and zero requests for embedded secrets.
promptingsecrets
42
Design a summarization prompt family for 14 document types and six languages, with 92% fact retention, 1,200-token outputs, and a $0.025 request cap.
promptingtokensdesign
43
Design an extraction prompt family for 35 schemas and 5 million pages per month, with 98.5% field accuracy and fewer than 0.3 repair calls per document.
promptingschemadesign
44
Design a classification prompt family for 180 intents, 10 million tickets per month, and English, German, and Japanese, with macro-F1 above 0.90.
classificationpromptingdesign
45
Set prompt-specific cost governance for 600 templates spending $1.8 million monthly, with a 20% reduction target and no quality slice losing more than one point.
prompting
46
A team of six prompt engineers proposes copying a 7,000-token prompt into 30 workflows to meet a six-week deadline. How would you mentor the lead while keeping delivery on schedule?
promptingtokensmentoring
47
Localize a 110-prompt healthcare fleet into Arabic and Japanese in eight weeks, with two native reviewers per language and critical safety recall above 99.5%.
prompting
48
Design prompts for resolving conflicts among up to 12 supplied policy excerpts, with 95% correct precedence decisions and explanations below 600 tokens.
promptingtokensdesign
49
Provide prompt-level audit evidence for 70 regulated templates reviewed by three external auditors, with evidence delivery in five business days and seven-year reproducibility.
prompting
50
Design a company-wide prompt architecture for 1,200 templates, 25 teams, six languages, and a $3 million annual prompt-evaluation budget, with 99.9% publishing availability.
promptingarchitecturedesign
51
A 14-line system prompt edit drops support resolution from 89% to 76% across 3,000 evaluations. The release is scheduled for 16:00 today. What do you do?
promptingsystem-designllm-eval
52
A migration moves 9 role instructions from user messages into the system prompt, and policy compliance falls from 97% to 83%. You have until tomorrow at 11:00 to decide whether to continue. How do you investigate?
promptingmigrationssystem-design
53
A provider update changes how negative instructions are followed, raising prohibited outputs from 0.4% to 3.1% on 8,000 tests. A go or no-go decision is due in 4 hours. What is your response?
testing
54
A customer name containing a closing XML tag escapes a prompt variable and overrides the summarization instruction in 62 of 20,000 requests. You must contain it within 90 minutes. What do you change?
promptingxml-sections
55
A document includes the same triple-delimiter used by the prompt template, causing 7.8% of 1,200 summaries to follow document instructions. The patch must ship by 18:00. How do you fix it?
prompting
56
A few-shot example copied from a real ticket contains a hidden instruction, and refund approvals rise from 6% to 19% in a 10,000-case replay. You have 3 hours before the next rollout. What do you do?
promptingfew-shot
57
Moving 6 few-shot examples reverses which answer wins in 24% of 2,400 pairwise tests. The experiment readout is tomorrow at 09:00. How do you handle the ordering bias?
promptingfew-shotsoft-skills
58
A chain-of-thought prompt scores 82% on a 5,000-item reasoning benchmark, and the signed target is 90% by Friday. How would you pursue the gap without exposing hidden reasoning?
promptingchain-of-thoughtbenchmarking
59
A debugging prompt exposes internal rationale and an API credential in 38 production responses. Security needs containment in 30 minutes and a fix by 17:00. What do you do?
promptingapi
60
A self-consistency prompt raises answer accuracy from 87% to 89% but increases monthly cost from $90,000 to $310,000. Finance needs a decision by noon tomorrow. What do you recommend?
promptingself-consistencyconsistency
61
A generate-then-critique prompt changes correct tax answers into wrong ones in 11% of 900 reviewed cases. The release window closes in 5 hours. How do you respond?
prompting
62
A ReAct prompt repeats the same malformed action 17 times and consumes $4.20 on one task with a $0.20 budget. You have 2 hours to stop recurrence. What do you change?
promptingreact
63
A planner prompt spends 70% of a 12,000-token budget on planning and leaves too little context to finish 31% of tasks. A revised prompt is due by 15:00. What is your design?
promptingtokenscontext
64
A tool description says it can update any account, and red-team prompts trigger 14 unauthorized update attempts in 500 tests. The tool must be safe by tomorrow at 10:00. What do you change?
promptingtestingred-teaming
65
A tool-argument prompt copies a user-supplied customer ID instead of the authenticated ID in 3.6% of 4,000 calls. You have until 13:00 to prevent cross-account requests. What do you do?
prompting
66
In a 4-stage prompt chain, the extractor renames confidence to score and the verifier silently accepts missing confidence in 22% of 6,000 runs. The fix is due in 6 hours. How do you handle it?
promptingsoft-skills
67
A prompt edit drops valid JSON from 99.6% to 91.2% across 80,000 daily requests. Operations needs recovery within 45 minutes. What is your plan?
promptingjson-output
68
A schema prompt treats a missing discount as 0 in English and null in Spanish, corrupting 8,700 records before a 19:00 reconciliation deadline. How do you resolve the nullable/default confusion?
promptingschemafundamentals
69
The correct policy appears first in the supplied context, but the answer prompt ignores it in 16% of 2,500 cases. A customer demo is in 7 hours. How do you fix the prompt?
promptingcontext
70
A citation prompt invents plausible section numbers in 9% of 1,600 legal answers. You have until tomorrow at 08:00 to make the workflow releasable. What do you change?
citationsprompting
71
Two supplied policies conflict, and the prompt chooses the older one in 29% of 1,000 tests despite visible dates. A decision is due by 14:00. How do you encode conflict and recency?
promptingtesting
72
A retrieved page says to ignore the system rules, and the answer prompt follows it in 23 of 400 red-team cases. You have 60 minutes to contain the issue. What do you do?
promptingsystem-designred-teaming
73
A direct jailbreak bypasses a safety prompt in 4.8% of 2,000 attempts and produces prohibited instructions. The public endpoint must be remediated within 2 hours. What is your response?
promptingendpointsllm-safety
74
Base64, zero-width characters, and mixed-script text bypass the injection prompt in 37 of 3,000 tests. You have until 17:30 to patch and validate it. What do you do?
promptinginjectionvalidation
75
Users reconstruct 86% of a confidential system prompt after 120 extraction attempts. Legal wants an exposure assessment by tomorrow at 10:00. How do you respond?
promptingsystem-designsystem-prompt
76
A constitutional prompt contains 12 rules, and rules 4 and 9 conflict, causing 18% inconsistent refusals across 2,200 cases. The policy review is in 6 hours. What do you do?
prompting
77
A safety prompt cuts unsafe answers from 2.1% to 0.2% but raises benign refusals from 5% to 28% on 7,500 cases. Product needs a release decision by 12:00. What do you recommend?
prompting
78
Email addresses and medical notes appear in 2.4% of rendered prompts and tracing logs for 18,000 requests. Privacy needs containment in 1 hour and scope by end of day. What do you do?
prompting
79
Eighteen tenants can insert custom prompt blocks, and those blocks weaken a mandatory base policy in 3.2% of 6,000 tests. A safe patch is due by 15:00. What do you do?
promptingtesting
80
An English prompt revision improves accuracy by 5 points, but German and Japanese fall by 12 and 17 points across 6,000 evaluations. Global rollout is planned in 8 hours. What do you decide?
promptingllm-eval
81
A translated safety prompt allows prohibited financial advice in 6.2% of Portuguese tests versus 0.5% in English. The locale launches tomorrow at 09:00. What do you do?
promptingtesting
82
At 96,000 input tokens, a critical instruction in the middle is missed in 21% of 1,500 cases, while 16,000-token cases pass. A mitigation is due by 18:00. What do you change?
tokens
83
A prompt compressor removes a single no-refund-without-receipt rule, raising unauthorized approvals from 1% to 13% in 4,000 tests. You have 2 hours to fix the release. What do you do?
promptingtesting
84
A system prompt grows from 1,800 to 7,400 tokens, raising p95 latency from 1.4 to 2.8 seconds and monthly cost by $170,000. A reduction plan is due Friday. How do you proceed?
promptingtokenslatency
85
Adding a request timestamp before the 2,200-token stable prefix drops cache hits from 81% to 9% and adds $95,000 monthly cost. The hotfix deadline is 15:00. What do you change?
tokenscachingestimation
86
The prompt registry serves version 41 to 8% of traffic even though version 43 is approved, causing a 10-point quality split. You have 1 hour to stabilize production. What do you do?
promptingregistries
87
An unversioned prompt hotfix improves one refund case but reduces overall policy accuracy from 94% to 81% across 5,000 replays. A formal replacement is due by 17:00. What do you do?
prompting
88
A rollback restores the system prompt but leaves 5 new few-shot examples active, so the error rate remains 8% instead of 1%. You have 45 minutes to complete recovery. What do you do?
promptingfew-shotsystem-design
89
Judge rubric prompt version 12 changes two scoring anchors, making 18 months of historical scores noncomparable and moving the release candidate by 6 points across 12,000 evaluations. A decision is due in 3 hours. What do you do?
promptingllm-eval
90
A pairwise judge chooses the first answer 64% of the time and prefers responses over 500 words by 18 points. You need a trustworthy result by tomorrow at noon. How do you fix the judge prompt?
prompting
91
Human reviewers prefer prompt A in 61% of 800 blinded pairs, while the judge prompt prefers B in 68%. A launch decision is due in 5 hours. Which signal do you use?
promptinghuman-eval
92
You discover that 320 of 2,000 golden-set answers were copied into few-shot examples, inflating prompt quality from 84% to 93%. The report is due tomorrow at 09:00. What do you do?
promptingfew-shot
93
A 9,000-case eval suite passes a prompt release, but complaints triple among voice users over 65 within 48 hours. A mitigation plan is due by 16:00. What do you do?
prompting
94
An A/B test shows a 4-point aggregate gain for prompt B, but Spanish billing accuracy falls from 92% to 71% across 1,100 cases. Full rollout is scheduled Friday. What do you decide?
promptingaggregationab-testing
95
Six weeks after launch, prompt success drifts from 90% to 82%, and 1,400 comments cluster around overly rigid answers to new request types. You need a diagnosis by Monday. What do you do?
promptingiac
96
You must migrate 27 production prompts to a new model in 21 days, and initial replay shows quality down 9 points with 14% more refusals. How do you run the migration?
promptingmigrations
97
A refund policy changes at 10:00, but an old prompt example still says customers have 60 days instead of 30, and 780 users receive the wrong answer by 13:00. The correction is due within 1 hour. What do you do?
prompting
98
During code review, you find a prompt change that concatenates raw user text into the system role and skips the 2,500-case injection suite to meet a 17:00 merge deadline. What do you do?
promptingcode-reviewestimation
99
A middle prompt engineer composes six prompt blocks in the wrong precedence, leaving two obsolete directives active and dropping claims accuracy from 93% to 79% across 26,000 requests. You have 24 hours to recover and mentor them. What do you do?
promptingmentoring
100
A 3,200-token legal disclaimer block cuts answer completion from 94% to 71% and shows no measurable safety gain, but launch is due in 2 hours. What do you recommend?
tokens

Questions

1Design an enterprise registry for 600 production prompt templates owned by 14 teams, with 99.9% availability and a 10-minute publication SLA.

2Define an immutable prompt bundle manifest for 180 templates, three providers, six languages, and a rollback target below five minutes.

3Design template composition and inheritance for 400 prompts across 12 products, while limiting any rendered prompt to 6,000 input tokens and two inheritance levels.

4Create role and provider adapters for 90 prompts that must behave consistently on GPT, Claude, and Gemini, with less than a two-point quality gap and a $0.03 request budget.

5Plan a 30-day migration of 250 customer-support prompts from GPT to Claude and Gemini across English, Spanish, and Japanese, with no more than 1% task-quality regression.

6Design a prompt compiler and linter for 1,000 templates changed by 20 teams, with CI feedback below 90 seconds and fewer than 0.5% false blocks.

7Specify variable typing and escaping for 300 templates receiving user text up to 20,000 characters in six languages, with zero instruction-boundary escapes in 10,000 adversarial tests.

8Design prompt and output contract schemas for a 24-field claims workflow processing 2 million documents per month with 99.95% contract validity.

9Architect a structured-output prompt fleet of 220 extractors serving 8 million calls per day, with 99.99% valid JSON and at most one repair attempt.

10Design a five-stage prompt chain for 500,000 contract reviews per month, with typed stage contracts, p95 below 12 seconds, and a $0.18 review budget.

11Separate planner and executor prompts for an agent limited to eight tool calls, 45 seconds, 12,000 total prompt tokens, and two replans.

12Write the architecture for ReAct policy prompts used by 40 workflows, with a maximum of six reasoning-action cycles and fewer than 0.2% policy-violating action proposals.

13Govern 75 tool descriptions shared by 10 agent teams, with a 4,000-token tool-description budget and 98% correct tool selection.

14Design answer, citation, and conflict prompts for a policy assistant handling 3 million questions per month, with 97% citation precision and support for documents from three authority levels.

15Order a 48,000-token prompt containing 30 evidence sections for a legal summarizer, with at least 90% fact retention and output below 2,000 tokens.

16Design a constitutional prompt hierarchy for 120 templates across four risk tiers, with policy updates published within 24 hours and zero unauthorized lower-level overrides.

17Build a safety prompt library for 16 product teams and eight harm categories, with critical-case recall above 99% in English and Spanish.

18Design an automated prompt-injection and red-team suite for 200 prompts, 25 attack families, six languages, and a four-hour nightly budget.

19Define tenant customization boundaries for 2,500 customers editing 140 base prompts, with 99.95% safe rendering and a 3,000-token customization cap.

20Operate a multilingual fleet of 160 prompts in English, Russian, French, German, Portuguese, and Spanish, with language quality within three points and releases inside 48 hours.

Locked questions