Exam-style
An insurer runs five customer-facing services on Amazon Bedrock, all sending the same system wording ahead of the customer's question so tone and refusal rules match wherever a customer lands. Somebody edited that wording last Tuesday. The answers have been vaguer since, two services no longer give a date for a claim, nobody kept the old text, and reverting means five deploys. What should the team put in place?
Reveal the answer
C. Put the wording in Amazon Bedrock Prompt Management as a prompt resource with input variables, a model and inference configuration, and numbered versions each service invokes by ARN
Amazon Bedrock Prompt Management holds a prompt as a resource rather than a string inside one application. The resource carries the message text with its input variables, the model it runs against, and the inference configuration, temperature and maximum output tokens included. Creating a version snapshots the working draft, numbered from 1 upwards; later edits to the draft leave it unchanged. Each service passes a prompt version ARN as the model identifier on a Converse call, with values for the variables. A restore is then a changed version number in five places, not five code changes and five releases. The internal library is the closest wrong option. Source control gives an author, a timestamp, a diff and a tag to revert to. The wording still ships inside the application, so a restore is a release in all five services, and the file records no model or inference settings. An Amazon S3 object overwritten in a bucket without versioning leaves no copy of the old wording. AWS Secrets Manager stores credentials and connection strings; a system prompt goes into the context window of every request. A guardrail filters input and output at invocation and restores nothing.
Q. Five services share one system prompt, last Tuesday’s edit made the answers worse, and reverting it means a deploy each. What goes in?
A. Amazon Bedrock Prompt Management: the wording becomes a prompt resource with input variables, an attached model and inference configuration, and numbered versions invoked by ARN.
Why? A restore is a version number changed in five places instead of five releases. The resource holds what shapes the output: the text with its variables, the model, and the inference settings it was tested at. A version snapshots the draft, and later edits leave it unchanged, so the version that passed review is the one still running. Pin a version in every environment with a customer on the other end. The near miss is keeping prompt templates in source control as a shared library. That is real versioning: authors, diffs, a tag to revert to. The restore is still a release in all five services, and the file records no model or temperature. An Amazon S3 object overwritten without bucket versioning leaves no copy. AWS Secrets Manager is for credentials, and a system prompt goes into the context window of every request. A guardrail filters input and output; it restores nothing.