Task 1.6: Implement prompt engineering strategies and governance for FM interactions.

• AStore the template as a JSON file in a versioned Amazon S3 bucket and have each application download the latest version at startup

  Incorrect

  S3 versioning keeps history, but there is no variant testing, and the applications still have to assemble the request because Converse cannot invoke a file in S3.

• BStore the template in an AWS Systems Manager Parameter Store parameter with a label for each approved wording

  Incorrect

  Parameter Store can hold text with labels, but it offers no variant comparison and cannot be invoked as a prompt by the Converse API.

• CKeep the template in code but generate it with a shared library that every application imports as a dependency

  Incorrect

  A shared library still ties every wording change to a build and deployment of each application.

• DUse Amazon Bedrock Prompt management with variables and variants, create versions, and call Converse with the version ARN

  Correct

  Prompt management stores templates with variables, lets you compare variants, creates immutable versions, and Converse accepts a prompt ARN as the model ID with promptVariables to fill the placeholders.

Prompt management treats prompts as managed assets: variables, variants for comparison, a draft and immutable versions, and direct invocation by ARN. Pinning applications to a version answers "which wording produced this output", and the invocation logs record the requests.

• APass the JSON schema in outputConfig.textFormat so that structured outputs constrain the response

  Correct

  Structured outputs constrain the model response to the supplied JSON schema, which removes parsing failures without custom retry logic.

• BAsk for XML instead of JSON in the system prompt because models follow XML tags more reliably

  Incorrect

  Changing the format does not change the fact that prompt-based formatting is not guaranteed, and the downstream service expects JSON.

• CLower the temperature to 0 and add a stop sequence after the closing brace of the JSON object

  Incorrect

  A low temperature makes output more deterministic but does not guarantee schema compliance, and a stop sequence only cuts the output off.

• DAdd a Lambda function that validates each response and calls the model again until the JSON parses

  Incorrect

  Validation and retry is the prompt-based approach that structured outputs replace. It adds latency and cost and still has a failure rate.

Prompting raises the odds of well-formed output; structured outputs guarantee it by constraining generation to a JSON schema. Remember the supported subset of JSON Schema Draft 2020-12: for example, minLength and numeric minimum or maximum are rejected with a 400 error.

• AUse Amazon EventBridge rules that route the classifier's output to targets based on event patterns

  Incorrect

  EventBridge routes events between services, but it does not hold the multi-step prompt chain, call the knowledge base, or provide versioned releases of the workflow.

• BCreate one long prompt in Prompt management that asks the model to classify, summarize or decline, and format the result

  Incorrect

  A single prompt cannot query the knowledge base or run the Lambda formatter, and it makes the branching unpredictable.

• CBuild an Amazon Bedrock flow with prompt, condition, knowledge base and Lambda nodes, and invoke it through an alias

  Correct

  Flows chain prompt nodes, condition nodes and knowledge base and Lambda function nodes in a visual builder. Applications call an alias that points to an immutable version, so releases and rollbacks are alias changes.

• DWrite a Lambda function that calls the classifier, the knowledge base and the formatter in sequence with if-else logic

  Incorrect

  A single function works but puts the branching in code, so analysts cannot maintain it visually and each change needs a deployment.

Prompt chains with conditional branches and pre- and post-processing are what Bedrock Flows is for. Prompt nodes can reuse Prompt management prompts, and versions with aliases give release and rollback without code changes.

• AAdd a role line such as "You are an expert email analyst"

  Incorrect

  A persona shapes tone and expertise, as in a support assistant with a set persona. It does not define which outputs are allowed.

• BList the eight category names as the answer choices at the end of the prompt

  Correct

  Guidance for classification is to name the choices explicitly, because without them the model tends to produce a free-form summary, and to place the choices at the end of the prompt.

• CAsk the model to explain its reasoning step by step before it classifies

  Incorrect

  Step-by-step reasoning helps with multi-step deductions. It adds text around the label that the router must match exactly.

• DAdd a stop sequence after the first line of the output

  Incorrect

  A stop sequence ends generation after a given string. A first line that is part of a summary is still not a category name.

Classification prompts work best when the allowed answers are named and placed where the model reads the task: at the end.

• AAt the end of the prompt, after the contract and the background

  Correct

  Putting the task or question at the end helps the model determine which information it has to find and stay focused on the task.

• BIn a knowledge base query that retrieves only the most similar clauses

  Incorrect

  Retrieval fits large or many documents. Reviewers need the whole contract considered, and it already fits in the context window.

• CAt the top as now, with maxTokens raised so that the model has room to answer

  Incorrect

  A larger response length helps when answers are cut off. The problem is what the model answers, not how much room it has.

• DAt the top, followed by an instruction to restate the question before answering

  Incorrect

  The question would still come before 40,000 tokens of other text, which is the placement that makes the model lose the task.

For open-book questions, put the document first and the question last.

• ALower maxTokens to 60 tokens for every summary request

  Incorrect

  A response length limit caps tokens, and the model stops wherever the cap falls, often mid-sentence. It also does nothing about the advice.

• BTell the model to summarize only what the ticket says, with no advice

  Correct

  Simple, clear and complete instructions reduce ambiguity about the task, including what the summary may contain.

• CFine-tune the model on 200 approved three-bullet summaries

  Incorrect

  Fine-tuning learns a task from labeled examples. It is the training project the team wants to avoid.

• DState the format: exactly three bullets of at most 20 words each

  Correct

  Output indicators describe constraints on the output, such as format and length, so the model produces the expected shape.

• EAdd a stop sequence after the third bullet marker

  Incorrect

  A stop sequence ends generation after a given string. It does not make the model write bullets or stay on topic.

Tell the model what to produce and from what: a scope instruction controls content, and an output indicator controls shape and length.

• AKeep each conversation in a variable declared outside the Lambda handler

  Incorrect

  A variable outside the handler exists only in one execution environment. Requests land in any of hundreds of environments that never see it, and Lambda replaces environments every few hours.

• BHave each client app send the full conversation history with every request

  Incorrect

  Client-held history works when one client holds the whole conversation. Here the other device lacks the earlier turns, and the history is not trimmed.

• CIngest each turn into a knowledge base and retrieve the most similar turns

  Incorrect

  A knowledge base suits retrieval over documents. New turns are searchable only after a sync, and similarity does not return the most recent turns in order.

• DStore turns in DynamoDB with a time sort key and TTL, and query the newest 20

  Correct

  Query returns a conversation's items sorted by sort key, and ScanIndexForward false with Limit 20 reads the newest turns. A TTL attribute lets DynamoDB delete expired items automatically.

Conversation context must live outside stateless compute. A keyed store with ordered reads and expiry gives every device and every execution environment the same bounded history.

• ASet temperature to 0 and lower top P on every request

  Incorrect

  Lower values make the model favor higher-probability tokens, so answers vary less. Nothing in the prompt changes, so the model still cannot tell an example from policy or this year's rules from last year's.

• BLoad both policies into a knowledge base and retrieve passages per question

  Incorrect

  A knowledge base finds passages relevant to each query, which suits large document sets. It does not stop the model from copying the examples or mark which policy is current, and leaving last year's policy out would break the comparison questions.

• CWrap each part in its own named markers and point the instructions at them

  Correct

  Clear delimiters, such as start and end markers or tags, help the model tell the parts of a prompt apart. The instructions can then name each part: answer from this year's policy, use last year's only to compare, and never repeat the examples.

• DFine-tune the model on approved answers drawn from the current policy

  Incorrect

  Fine-tuning adjusts model parameters with labeled training data. Every policy revision would need new data and a new training job, which does not fit a fix this week, and it does not label the texts in the prompt.

When one prompt carries several kinds of text, delimit each part and refer to the parts by name in the instructions.