Topic 04 Practice Test 1 covers Prompt Engineering, Prompt Management, Context, Prompt Flows, and Governance in Domain 1 (Task 1.6). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
A healthcare assistant must always identify itself as a support tool, answer in a fixed three-section format, and keep those instructions separate from user-provided text. Which implementation best establishes that behavior? Choose ONE.
- Increase maxTokens for the prompt to change the maximum generated-response length
- Put the role and response contract in the system prompt
- Store the conversation in Amazon DynamoDB to persist application state across requests
- Append the instructions to every user message as a separate configuration change
Correct Answer: B
Correct Answer
Answer B is correct because Put the role and response contract in the system prompt provides the needed capability. A system prompt supplies high-priority instructions and persistent context to supported Converse-capable models. For this workload, This directly separates application-level behavior from changing user content and can carry the required identity and output contract. Constraint met: application instructions must be separate from user text.
Incorrect Answers
Answer A is incorrect because maxTokens caps how many tokens the model may generate in its response. Increase maxTokens for the prompt to change the maximum generated-response length can help elsewhere, but here A larger generation ceiling does not establish role identity, section ordering, or instruction precedence, so it cannot enforce the required behavior. Unmet condition: application instructions must be separate from user text.
Answer C is incorrect because Store the conversation in Amazon DynamoDB to persist application state across requests does not solve this control. DynamoDB can persist conversation state across requests or sessions. In this scenario, Persistence can support continuity, but stored history does not itself define the assistant role or a mandatory response structure. It fails: application instructions must be separate from user text.
Answer D is incorrect because Append the instructions to every user message as a separate configuration change serves another purpose: User-message text is ordinary conversational input and can be mixed with the requester’s own content. Here, Repeating policy in each user turn weakens separation between application instructions and user-supplied material, which is the key governance requirement. It misses this constraint: application instructions must be separate from user text.
Question 2
A customer-support application must prevent account numbers and email addresses from appearing in model responses while still allowing the rest of an answer to be returned. Which control best meets the requirement? Choose ONE.
- Create a second prompt variant to compare an alternative prompt configuration
- Use a Guardrails denied-topic policy to apply a configured model-interaction safeguard
- Set temperature to zero to change sampling randomness during generation
- Configure an Amazon Bedrock Guardrails sensitive-information filter to mask detected PII
Correct Answer: D
Correct Answer
Answer D is correct because Configure an Amazon Bedrock Guardrails sensitive-information filter to mask detected PII is appropriate because sensitive-information filters can detect supported PII entities or custom regex patterns and can mask sensitive values in prompts or responses. Here, Masking preserves the permitted portions of the response while removing the account and email values that the policy forbids exposing. This satisfies: response may contain account numbers and email addresses.
Incorrect Answers
Answer A is incorrect because Create a second prompt variant to compare an alternative prompt configuration is adjacent but not sufficient. Prompt variants let teams compare alternative prompt text, models, or inference configurations. Here, A variant is useful for experimentation, not deterministic privacy enforcement on every input and output. The unmet need is: response may contain account numbers and email addresses.
Answer B is incorrect because Denied topics block content associated with defined subject areas. Use a Guardrails denied-topic policy to apply a configured model-interaction safeguard can help elsewhere, but here The problem is not a prohibited subject; it is sensitive values embedded in otherwise valid support content, so topic blocking is too broad and misaligned. Unmet condition: response may contain account numbers and email addresses.
Answer C is incorrect because Set temperature to zero to change sampling randomness during generation has a different function. Temperature influences randomness in token selection during generation. Under these conditions, Lower variability can improve repeatability, but it neither detects PII nor masks sensitive values after they appear. It does not meet: response may contain account numbers and email addresses.
Question 3
A travel assistant must remember a user’s preferred airport and loyalty tier after the user closes the browser and returns the next day. Which design best supplies durable conversational context? Choose ONE.
- Persist the relevant conversation state in DynamoDB and reload it for the next session
- Rely only on the current model context window to keep prior messages inside the model context
- Raise the prompt temperature to change sampling randomness during generation
- Publish a new Prompt Management version as a separate workflow technique
Correct Answer: A
Correct Answer
Answer A is correct because Persist the relevant conversation state in DynamoDB and reload it for the next session: DynamoDB provides durable key-value or document storage that an application can use for conversation history and user context. Here, Reloading the user-specific state after a new session satisfies the cross-session persistence requirement without depending on an in-memory chat process. This meets the case constraint: browser session ends.
Incorrect Answers
Answer B is incorrect because Rely only on the current model context window to keep prior messages inside the model context is adjacent but not sufficient. The model context window contains only the messages supplied with the current inference request. Here, Once the browser session and application state are gone, the model does not automatically recover yesterday’s preferences, so current-request context is insufficient. The unmet need is: browser session ends.
Answer C is incorrect because Temperature affects output diversity rather than persistence. Raise the prompt temperature to change sampling randomness during generation can help elsewhere, but here Changing sampling behavior does not store the airport or loyalty tier and therefore cannot restore context on a later day. Unmet condition: browser session ends.
Answer D is incorrect because Publish a new Prompt Management version as a separate workflow technique has a different function. Prompt versions snapshot reusable prompt configuration for deployment. Under these conditions, Versioning governs prompt definitions, not per-user memory, so it cannot retain an individual traveler’s prior preferences. It does not meet: browser session ends.
Question 4
An internal assistant receives requests such as “prepare the report,” but several report types exist. The team wants the assistant to ask a targeted clarification before continuing rather than guess. Which architecture best supports this interaction? Choose ONE.
- Store the prompt template in Amazon S3 to retain governed artifacts in object storage
- Use a higher topP value to change nucleus-sampling diversity during generation
- Use a Step Functions workflow that branches to a clarification step when required information is missing
- Add more stop sequences to the model configuration as a separate workflow technique
Correct Answer: C
Correct Answer
Answer C is correct because Use a Step Functions workflow that branches to a clarification step when required information is missing matches the need. Step Functions can coordinate deterministic states and branches around model interactions, including pausing a workflow until required inputs are collected. Specifically, The workflow can detect that report type is absent, request that specific detail, and resume only after the user supplies it. Deciding constraint: request lacks report type.
Incorrect Answers
Answer A is incorrect because Store the prompt template in Amazon S3 to retain governed artifacts in object storage does not solve this control. S3 can provide durable storage for prompt assets or repositories. In this scenario, Template storage improves governance and retrieval of artifacts, but it does not orchestrate a follow-up question based on missing information. It fails: request lacks report type.
Answer B is incorrect because Use a higher topP value to change nucleus-sampling diversity during generation serves another purpose: topP changes the candidate-token distribution used during generation. Here, More sampling diversity makes guessing less predictable; it does not enforce a deterministic clarification step before processing. It misses this constraint: request lacks report type.
Answer D is incorrect because Stop sequences terminate generation when specified text is emitted. Add more stop sequences to the model configuration as a separate workflow technique can help elsewhere, but here They can bound model output, but they do not create a stateful clarification branch or collect missing business parameters. Unmet condition: request lacks report type.
Question 5
One approved prompt must be reused for hundreds of products while product_name, locale, and policy_text change at runtime. Which Prompt Management feature should the team use? Choose ONE.
- Increase the flow maxIterations value as a separate configuration change
- Define prompt variables such as {{product_name}} and supply values at invocation time
- Create a separate prompt version for every product to create an immutable governed prompt snapshot
- Create one Guardrails denied topic for every product to refuse requests about a configured subject
Correct Answer: B
Correct Answer
Answer B is correct because Define prompt variables such as {{product_name}} and supply values at invocation time is appropriate because prompt Management variables are placeholders in a reusable prompt whose values are supplied during testing or runtime invocation. Here, Variables keep one governed template while safely substituting the changing product, locale, and policy data for each request. This satisfies: same approved template.
Incorrect Answers
Answer A is incorrect because Increase the flow maxIterations value as a separate configuration change has a different function. maxIterations limits how many times a DoWhile flow loop can execute. Under these conditions, Loop capacity has no relationship to substituting dynamic fields into a reusable prompt template. It does not meet: same approved template.
Answer C is incorrect because Create a separate prompt version for every product to create an immutable governed prompt snapshot serves another purpose: Versions are point-in-time snapshots of a prompt configuration intended for deployment and rollback. Here, Creating hundreds of nearly identical versions would misuse versioning and make governance harder when only runtime data differs. It misses this constraint: same approved template.
Answer D is incorrect because Create one Guardrails denied topic for every product to refuse requests about a configured subject is adjacent but not sufficient. Denied topics are safety controls that block configured subject areas. Here, They do not parameterize reusable prompt text and would not inject product-specific values into the model request. The unmet need is: same approved template.
Question 6
A prompt team is satisfied with a tested draft and wants production to use a stable snapshot while prompt authors continue editing the working draft. What should the team do? Choose ONE.
- Create a Prompt Management version from the approved draft and deploy that version
- Let production invoke the mutable draft directly as a separate configuration change
- Store only the prompt text in CloudWatch Logs to retain searchable operational log events
- Increase the prompt maxTokens setting to change the maximum generated-response length
Correct Answer: A
Correct Answer
Answer A is correct because A Prompt Management version is a point-in-time snapshot created from the working draft for use by applications. In this case, Create a Prompt Management version from the approved draft and deploy that version fits because Production can remain pinned to the approved snapshot while authors safely continue iterating on the mutable draft for future releases. Required condition: production needs stable approved prompt.
Incorrect Answers
Answer B is incorrect because Let production invoke the mutable draft directly as a separate configuration change has a different function. The draft is the working copy that prompt authors can continue modifying. Under these conditions, Using it in production would couple live behavior to ongoing edits, violating the requirement for a stable approved snapshot. It does not meet: production needs stable approved prompt.
Answer C is incorrect because Store only the prompt text in CloudWatch Logs to retain searchable operational log events does not solve this control. CloudWatch Logs provides operational log storage rather than prompt release management. In this scenario, Logs can provide evidence, but they neither create an immutable prompt snapshot nor give the application a deployable prompt version. It fails: production needs stable approved prompt.
Answer D is incorrect because Increase the prompt maxTokens setting to change the maximum generated-response length serves another purpose: maxTokens limits response length for inference. Here, Changing generation length does not establish release immutability or separate production from ongoing prompt edits. It misses this constraint: production needs stable approved prompt.
Question 7
Auditors need an API-level record showing who created, updated, or deleted governed prompt resources. Which AWS service is the most appropriate source of that management activity? Choose ONE.
- Amazon DynamoDB Streams to persist application state across requests
- Amazon Bedrock prompt variables to substitute runtime values into the template
- Amazon CloudWatch metrics as a separate workflow technique
- AWS CloudTrail
Correct Answer: D
Correct Answer
Answer D is correct because AWS CloudTrail is appropriate because cloudTrail records supported AWS API activity and identifies the principal, operation, time, and request context for governance investigations. Here, That event history is the right evidence for attributing prompt-management changes to specific identities and API actions. This satisfies: need identity and API action attribution.
Incorrect Answers
Answer A is incorrect because Amazon DynamoDB Streams to persist application state across requests is adjacent but not sufficient. DynamoDB Streams captures item-level changes from a DynamoDB table. Here, It can audit data mutations in that table, but prompt resources are not governed by a DynamoDB stream and the requirement is AWS API activity. The unmet need is: need identity and API action attribution.
Answer B is incorrect because Prompt variables substitute runtime values into prompt templates. Amazon Bedrock prompt variables to substitute runtime values into the template can help elsewhere, but here They improve reuse but do not record which identity changed a prompt resource or when that management action occurred. Unmet condition: need identity and API action attribution.
Answer C is incorrect because Amazon CloudWatch metrics as a separate workflow technique has a different function. CloudWatch metrics store numerical time-series data used for monitoring thresholds and trends. Under these conditions, Metrics can show rates or failures but do not provide the authoritative caller-and-API event record required for change attribution. It does not meet: need identity and API action attribution.
Question 8
A platform team wants application operators to search prompt invocation diagnostics, access records, and structured error messages without changing the prompt itself. Which service best fits this operational logging need? Choose ONE.
- A Bedrock Guardrails word filter to match configured words or phrases
- AWS CloudTrail only to record aws api activity for audit
- Amazon CloudWatch Logs
- Prompt Management variables as a separate configuration change
Correct Answer: C
Correct Answer
Answer C is correct because CloudWatch Logs centralizes application and service log events for search, retention, metric filters, and troubleshooting. In this case, Amazon CloudWatch Logs fits because It fits the need for searchable runtime access and error evidence while keeping prompt definition and deployment concerns separate. Required condition: operators need searchable runtime logs.
Incorrect Answers
Answer A is incorrect because A Bedrock Guardrails word filter to match configured words or phrases serves another purpose: Word filters block configured exact words or phrases in interactions. Here, Filtering user or model text does not capture application access records or operational errors for later analysis. It misses this constraint: operators need searchable runtime logs.
Answer B is incorrect because AWS CloudTrail only to record aws api activity for audit is adjacent but not sufficient. CloudTrail is strongest for AWS API governance and identity attribution. Here, It may show service API calls, but it is not a substitute for the application’s detailed runtime diagnostic and structured error log stream. The unmet need is: operators need searchable runtime logs.
Answer D is incorrect because Prompt Management variables as a separate configuration change has a different function. Variables provide placeholders for dynamic prompt content. Under these conditions, They have no logging or retention function and cannot make runtime diagnostics searchable by operators. It does not meet: operators need searchable runtime logs.
Question 9
A team has two prompt variants for summarizing legal documents and wants evidence about which one produces better outputs before promotion. What is the best first step? Choose ONE.
- Test both variants with representative inputs and compare their outputs in Prompt Management
- Create production versions of both variants immediately as a different governance mechanism
- Increase topP on both variants to change nucleus-sampling diversity during generation
- Store both variants as unrelated S3 objects to retain governed artifacts in object storage
Correct Answer: A
Correct Answer
Answer A is correct because Test both variants with representative inputs and compare their outputs in Prompt Management is appropriate because prompt Management supports running test values through different prompt variants so teams can compare behavior before selecting a preferred configuration. Here, Using the same representative cases creates direct evidence about output quality instead of promoting a variant based on author preference. This satisfies: two candidate prompt variants.
Incorrect Answers
Answer B is incorrect because Create production versions of both variants immediately as a different governance mechanism serves another purpose: Versions are appropriate after a configuration is considered ready for deployment. Here, Publishing both before comparative testing skips the quality gate the team specifically needs and introduces unnecessary production candidates. It misses this constraint: two candidate prompt variants.
Answer C is incorrect because Increase topP on both variants to change nucleus-sampling diversity during generation is adjacent but not sufficient. topP changes sampling behavior by limiting the cumulative probability mass considered for token selection. Here, Applying the same inference change does not establish which existing prompt design performs better on the required legal examples. The unmet need is: two candidate prompt variants.
Answer D is incorrect because S3 can archive prompt artifacts, but object storage does not execute the prompts or compare model outputs. Store both variants as unrelated S3 objects to retain governed artifacts in object storage can help elsewhere, but here The team needs behavioral evidence, not merely a second copy of each template. Unmet condition: two candidate prompt variants.
Question 10
A model must return JSON with fields id, risk, and rationale. The team wants an automated gate that rejects a prompt revision when required fields are missing. Which approach best provides that gate? Choose ONE.
- Create an S3 lifecycle rule for prompt artifacts to retain governed artifacts in object storage
- Use a word filter for the field names to match configured words or phrases
- Raise the model temperature during testing to change sampling randomness during generation
- Invoke a Lambda validator that parses the response and fails when the required schema is absent
Correct Answer: D
Correct Answer
Answer D is correct because Invoke a Lambda validator that parses the response and fails when the required schema is absent matches the need. A Lambda function can deterministically parse generated output, validate required fields and types, and return a pass or fail result to the surrounding workflow. Specifically, Schema validation is objective and machine-checkable, so it gives the prompt revision a repeatable acceptance gate. Deciding constraint: required fields id risk rationale.
Incorrect Answers
Answer A is incorrect because Create an S3 lifecycle rule for prompt artifacts to retain governed artifacts in object storage has a different function. S3 lifecycle rules manage object retention and storage transitions. Under these conditions, Retention policy does not inspect model responses or fail a prompt revision when its output schema is incorrect. It does not meet: required fields id risk rationale.
Answer B is incorrect because Use a word filter for the field names to match configured words or phrases does not solve this control. Guardrails word filters block configured words or phrases. In this scenario, Blocking the required JSON keys would work against the desired output and still would not validate object structure or data types. It fails: required fields id risk rationale.
Answer C is incorrect because Raise the model temperature during testing to change sampling randomness during generation serves another purpose: Higher temperature generally increases output diversity. Here, That may expose variability, but it cannot deterministically determine whether every required JSON field is present and valid. It misses this constraint: required fields id risk rationale.
Question 11
Before releasing a new customer-service prompt, QA must exercise normal requests, adversarial wording, missing fields, and timeout paths in a controlled sequence. Which orchestration best fits? Choose ONE.
- Enable a Guardrails denied topic to refuse requests about a configured subject
- Use Step Functions to run the edge-case test workflow and record each branch result
- Use only a larger context window to keep prior messages inside the model context
- Create a new prompt variable named test to substitute runtime values into the template
Correct Answer: B
Correct Answer
Answer B is correct because Use Step Functions to run the edge-case test workflow and record each branch result provides the needed capability. Step Functions coordinates multi-step workflows, explicit branches, retries, and failure paths around service calls. For this workload, It can execute the planned normal, adversarial, missing-input, and timeout cases consistently and preserve the outcome of each branch. Constraint met: must exercise multiple paths in controlled sequence.
Incorrect Answers
Answer A is incorrect because Denied topics block interactions about configured subjects. Enable a Guardrails denied topic to refuse requests about a configured subject can help elsewhere, but here This can be one safety control under test, but it is not a general harness for executing the complete release test matrix. Unmet condition: must exercise multiple paths in controlled sequence.
Answer C is incorrect because Use only a larger context window to keep prior messages inside the model context does not solve this control. More available context can hold additional messages or reference text. In this scenario, Context capacity does not schedule test cases, branch on failures, or ensure that all required edge conditions are exercised. It fails: must exercise multiple paths in controlled sequence.
Answer D is incorrect because Create a new prompt variable named test to substitute runtime values into the template serves another purpose: A prompt variable supplies runtime content to a template. Here, One placeholder can carry data but cannot orchestrate the sequence, branching, retry, and result capture required by the QA plan. It misses this constraint: must exercise multiple paths in controlled sequence.
Question 12
After a prompt release, the team needs an alert when the percentage of responses failing its validator rises above the historical baseline. Which design best detects that regression? Choose ONE.
- Create a new Prompt Management variable for failure_rate as a different governance mechanism
- Use CloudTrail to compare every model response to record aws api activity for audit
- Publish validator results as CloudWatch metrics and alarm on the failure-rate threshold
- Increase maxTokens whenever a validation fails to change the maximum generated-response length
Correct Answer: C
Correct Answer
Answer C is correct because Publish validator results as CloudWatch metrics and alarm on the failure-rate threshold provides the needed capability. CloudWatch metrics and alarms can track a numerical quality signal over time and notify when it breaches an expected threshold. For this workload, A validator-failure percentage is exactly the kind of regression indicator that can be aggregated and alarmed after deployment. Constraint met: post-release quality regression.
Incorrect Answers
Answer A is incorrect because Create a new Prompt Management variable for failure_rate as a different governance mechanism is adjacent but not sufficient. Prompt variables inject values into prompt text at invocation time. Here, They do not calculate observed failures across requests or raise monitoring alarms when quality deteriorates. The unmet need is: post-release quality regression.
Answer B is incorrect because CloudTrail records AWS API activity rather than the semantic quality of each generated payload. Use CloudTrail to compare every model response to record aws api activity for audit can help elsewhere, but here It is useful for governance evidence, but response-validation outcomes should be emitted as application monitoring data. Unmet condition: post-release quality regression.
Answer D is incorrect because Increase maxTokens whenever a validation fails to change the maximum generated-response length does not solve this control. maxTokens changes the permitted response length. In this scenario, Automatically expanding output length is unrelated to measuring a fleet-wide failure percentage and could introduce new variability. It fails: post-release quality regression.
Question 13
A classification prompt produces correct labels but occasionally adds prose that breaks a downstream parser. The application needs only one label from a fixed set. Which prompt change most directly improves reliability? Choose ONE.
- Store the conversation history in DynamoDB to persist application state across requests
- Add a flow iterator node to process array elements through repeated steps
- Raise temperature to increase creativity to change sampling randomness during generation
- Explicitly require exactly one allowed label and no additional text
Correct Answer: D
Correct Answer
Answer D is correct because Explicitly require exactly one allowed label and no additional text provides the needed capability. A strict output-format instruction narrows the response contract and tells the model what syntax is acceptable. For this workload, The failure is unwanted prose rather than missing knowledge, so specifying the exact label-only format directly targets the parser issue. Constraint met: downstream parser accepts only fixed labels.
Incorrect Answers
Answer A is incorrect because Store the conversation history in DynamoDB to persist application state across requests serves another purpose: DynamoDB can preserve prior turns and user state. Here, Persistent history does not constrain the current response to the allowed labels and may add irrelevant context to a simple classifier. It misses this constraint: downstream parser accepts only fixed labels.
Answer B is incorrect because Add a flow iterator node to process array elements through repeated steps is adjacent but not sufficient. An iterator processes array items one by one in a Bedrock Flow. Here, Iteration changes how a list is processed, not the format contract of a single classification response. The unmet need is: downstream parser accepts only fixed labels.
Answer C is incorrect because Higher temperature usually increases sampling variability. Raise temperature to increase creativity to change sampling randomness during generation can help elsewhere, but here Creativity is counterproductive when the downstream system needs a deterministic, machine-readable label with no extra explanation. Unmet condition: downstream parser accepts only fixed labels.
Question 14
A deterministic extraction task sometimes returns different field choices for identical inputs. Accuracy is already acceptable, and the team wants less sampling variability without rewriting the prompt. What should it change first? Choose ONE.
- Add an S3 storage node after inference
- Lower the temperature
- Increase topP toward 1.0 to change nucleus-sampling diversity during generation
- Create more prompt variables to substitute runtime values into the template
Correct Answer: B
Correct Answer
Answer B is correct because Lower the temperature: Temperature controls how strongly generation favors higher-probability token choices; lower values generally reduce randomness. Here, For an already-correct extraction prompt, decreasing sampling variability is the most direct configuration change to make repeated outputs more consistent. This meets the case constraint: same inputs should yield more consistent field choices.
Incorrect Answers
Answer A is incorrect because Add an S3 storage node after inference does not solve this control. An S3 storage node can persist flow output as an object. In this scenario, Saving the response may aid auditing, but storage occurs after generation and cannot reduce variability in the generated field choices. It fails: same inputs should yield more consistent field choices.
Answer C is incorrect because Increase topP toward 1.0 to change nucleus-sampling diversity during generation is adjacent but not sufficient. A larger topP lets the model consider a broader cumulative set of token candidates. Here, That generally preserves or increases diversity rather than tightening deterministic behavior for identical extraction inputs. The unmet need is: same inputs should yield more consistent field choices.
Answer D is incorrect because Variables let runtime values be substituted into a reusable prompt template. Create more prompt variables to substitute runtime values into the template can help elsewhere, but here Additional placeholders do not reduce stochastic sampling when the same prompt and input are invoked repeatedly. Unmet condition: same inputs should yield more consistent field choices.
Question 15
A model often misclassifies two similar support-ticket categories. The team has several short, approved examples that show the boundary between the categories. Which prompt technique is most appropriate? Choose ONE.
- Add a small set of representative labeled examples to the prompt
- Increase maxTokens to change the maximum generated-response length
- Move the prompt text to S3 to retain governed artifacts in object storage
- Use a Guardrails denied topic for one category to refuse requests about a configured subject
Correct Answer: A
Correct Answer
Answer A is correct because Add a small set of representative labeled examples to the prompt provides the needed capability. Few-shot examples demonstrate the desired input-to-output mapping and can clarify subtle class boundaries without modifying model weights. For this workload, The approved examples directly teach how near-neighbor tickets should be distinguished while keeping the task and output labels explicit. Constraint met: two legitimate categories are frequently confused.
Incorrect Answers
Answer B is incorrect because Increase maxTokens to change the maximum generated-response length does not solve this control. maxTokens raises or lowers the maximum response length. In this scenario, The classifier already needs a short label, so a larger output budget does not teach the distinction between the two confusing categories. It fails: two legitimate categories are frequently confused.
Answer C is incorrect because Move the prompt text to S3 to retain governed artifacts in object storage serves another purpose: S3 can store a governed copy of the template. Here, Repository location does not change the inference examples presented to the model and therefore does not improve the ambiguous class boundary. It misses this constraint: two legitimate categories are frequently confused.
Answer D is incorrect because Use a Guardrails denied topic for one category to refuse requests about a configured subject is adjacent but not sufficient. Denied topics are designed to block disallowed subject matter. Here, Both ticket categories are legitimate outputs, so blocking one would enforce the wrong behavior instead of improving classification precision. The unmet need is: two legitimate categories are frequently confused.
Question 16
A prompt is deployed to draft customer replies. Reviewers consistently edit overly formal greetings and add a missing escalation statement. What process best uses that evidence to improve the prompt? Choose ONE.
- Create a new DynamoDB table for every prompt revision to persist application state across requests
- Disable prompt versioning during experimentation to create an immutable governed prompt snapshot
- Capture reviewer feedback, update the prompt instructions, retest representative cases, and compare the new results
- Increase topP until reviewers stop editing to change nucleus-sampling diversity during generation
Correct Answer: C
Correct Answer
Answer C is correct because Capture reviewer feedback, update the prompt instructions, retest representative cases, and compare the new results is appropriate because a feedback loop turns observed quality defects into prompt changes and validates the revision against representative examples before promotion. Here, The recurring edits provide specific evidence about tone and escalation requirements, so iterative refinement directly addresses both defects. This satisfies: consistent tone and escalation edits.
Incorrect Answers
Answer A is incorrect because DynamoDB can store application state, but separate tables do not improve prompt behavior. Create a new DynamoDB table for every prompt revision to persist application state across requests can help elsewhere, but here The quality signal comes from reviewer edits and should drive instruction changes and regression testing, not storage proliferation. Unmet condition: consistent tone and escalation edits.
Answer B is incorrect because Disable prompt versioning during experimentation to create an immutable governed prompt snapshot has a different function. Versioning gives teams snapshots for controlled deployment and rollback. Under these conditions, Removing that governance mechanism makes iterative changes harder to compare and does nothing to incorporate the reviewer’s concrete feedback. It does not meet: consistent tone and escalation edits.
Answer D is incorrect because Increase topP until reviewers stop editing to change nucleus-sampling diversity during generation serves another purpose: topP controls token sampling diversity rather than encoding the missing escalation rule. Here, Changing randomness without translating reviewer feedback into explicit instructions is unlikely to fix a consistent policy omission. It misses this constraint: consistent tone and escalation edits.
Question 17
A Bedrock Flow receives a model-generated confidence score. Scores below 0.70 must go to a human-review path; all others continue automatically. Which flow component should make this routing decision? Choose ONE.
- A prompt variable to substitute runtime values into the template
- A condition node
- An iterator node as a separate workflow technique
- An S3 storage node to retain governed artifacts in object storage
Correct Answer: B
Correct Answer
Answer B is correct because A Bedrock Flow condition node evaluates relational or logical expressions and routes data through the matching conditional connection. In this case, A condition node fits because Comparing confidence with 0.70 and selecting one of two downstream paths is a direct condition-node use case. Required condition: branch on confidence threshold 0.70.
Incorrect Answers
Answer A is incorrect because A prompt variable to substitute runtime values into the template is adjacent but not sufficient. A prompt variable substitutes a supplied value into prompt text. Here, It can carry a score into a prompt, but it does not implement deterministic branch routing inside the flow graph. The unmet need is: branch on confidence threshold 0.70.
Answer C is incorrect because An iterator node as a separate workflow technique has a different function. An iterator emits items from an input array one at a time for downstream processing. Under these conditions, The requirement is a scalar threshold branch, not repeated processing of array elements, so iteration adds the wrong control structure. It does not meet: branch on confidence threshold 0.70.
Answer D is incorrect because An S3 storage node to retain governed artifacts in object storage does not solve this control. An S3 storage node writes flow data to an S3 object and returns its URI. In this scenario, Storage may preserve review evidence, but it cannot choose the human or automatic path based on the score. It fails: branch on confidence threshold 0.70.
Question 18
A team already manages an approved prompt in Prompt Management and wants a Bedrock Flow to reuse that governed prompt instead of copying its text inline. What should the prompt node reference? Choose ONE.
- The CloudWatch log group ARN as a separate workflow technique
- The DynamoDB table ARN storing session state
- The Guardrails denied-topic name to apply a configured model-interaction safeguard
- The managed prompt ARN
Correct Answer: D
Correct Answer
Answer D is correct because The managed prompt ARN provides the needed capability. A flow prompt node can use a resource source configuration that references a Prompt Management prompt by ARN. For this workload, Referencing the managed resource keeps the flow connected to the governed prompt artifact and avoids maintaining a duplicate inline definition. Constraint met: reuse approved Prompt Management artifact.
Incorrect Answers
Answer A is incorrect because The CloudWatch log group ARN as a separate workflow technique serves another purpose: A log group ARN identifies a destination for operational log events. Here, It contains no prompt template or inference configuration for the flow’s prompt node to execute. It misses this constraint: reuse approved Prompt Management artifact.
Answer B is incorrect because The DynamoDB table ARN storing session state is adjacent but not sufficient. A DynamoDB table can hold application context or conversation history. Here, The table may provide runtime data, but it is not a Prompt Management resource and cannot replace the prompt node’s prompt source. The unmet need is: reuse approved Prompt Management artifact.
Answer C is incorrect because A denied-topic name identifies one safety policy element inside a guardrail. The Guardrails denied-topic name to apply a configured model-interaction safeguard can help elsewhere, but here It does not identify the reusable prompt resource that the node must invoke. Unmet condition: reuse approved Prompt Management artifact.
Question 19
A production Bedrock Flow must remain on an approved workflow snapshot while developers keep changing and testing the draft. The team also needs a quick rollback path. Which deployment pattern best fits? Choose ONE.
- Store a copy of the flow definition in CloudWatch Logs to retain searchable operational log events
- Increase the condition-node count as a different governance mechanism
- Publish an immutable flow version and invoke it through an alias that can be repointed for rollback
- Invoke the working draft directly from production as a different governance mechanism
Correct Answer: C
Correct Answer
Answer C is correct because Publish an immutable flow version and invoke it through an alias that can be repointed for rollback matches the need. Publishing creates an immutable snapshot of the prepared flow, and an alias is the application-facing target used for InvokeFlow. Specifically, The alias can be moved between approved versions without exposing the mutable draft, which supports both controlled rollout and rollback. Deciding constraint: production must stay on approved snapshot.
Incorrect Answers
Answer A is incorrect because Store a copy of the flow definition in CloudWatch Logs to retain searchable operational log events does not solve this control. Logs are useful for diagnostics and evidence. In this scenario, A logged definition is not a deployable Bedrock Flow version and cannot receive InvokeFlow traffic or provide alias-based rollback. It fails: production must stay on approved snapshot.
Answer B is incorrect because Increase the condition-node count as a different governance mechanism serves another purpose: Condition nodes add routing logic inside a flow. Here, More branches do not create a release snapshot or an application-level pointer that can be changed during rollback. It misses this constraint: production must stay on approved snapshot.
Answer D is incorrect because The working draft is intended for iterative editing and testing. Invoke the working draft directly from production as a different governance mechanism can help elsewhere, but here Production would then inherit in-progress changes and lose the stable release boundary the team requires. Unmet condition: production must stay on approved snapshot.
Question 20
A long-running support chat is approaching the model context limit. The application must preserve the active issue, the last few turns, and durable customer preferences without resending every historic message. Which approach is best? Choose ONE.
- Summarize older turns, retain the recent exchange, and reload durable preferences from the conversation store
- Publish a new flow version after each user turn to preserve a published immutable workflow snapshot
- Always resend the entire raw conversation as a separate workflow technique
- Increase temperature as the chat grows to change sampling randomness during generation
Correct Answer: A
Correct Answer
Answer A is correct because Summarize older turns, retain the recent exchange, and reload durable preferences from the conversation store provides the needed capability. Context management can combine persisted state with a compact summary and the most relevant recent messages before inference. For this workload, This preserves decision-relevant history while controlling token growth, instead of blindly replaying every turn into the model context. Constraint met: approaching context limit.
Incorrect Answers
Answer B is incorrect because Publish a new flow version after each user turn to preserve a published immutable workflow snapshot does not solve this control. Flow versions snapshot workflow definitions for deployment. In this scenario, Per-turn version creation would confuse release governance with user state and would not compress or select conversation context. It fails: approaching context limit.
Answer C is incorrect because Always resend the entire raw conversation as a separate workflow technique serves another purpose: Supplying full history can preserve detail while the context window still has capacity. Here, The stated problem is approaching that limit, so unbounded replay will eventually fail or crowd out the current issue and instructions. It misses this constraint: approaching context limit.
Answer D is incorrect because Increase temperature as the chat grows to change sampling randomness during generation is adjacent but not sufficient. Temperature changes generation randomness. Here, It neither reduces input-token usage nor decides which historical facts should remain available to the model. The unmet need is: approaching context limit.
