Build a careful AI helper — Bedrock

Your goal

Plan an AI helper that drafts answers for people to check.

Why this matters

An AI model can write useful text, but it can also be wrong. You will call it through a protected server, send only suitable data, and keep a person in charge of important actions.

Words you will use

A model generates output from input. A prompt is the request you send. Bedrock provides managed access to foundation models.

How to proceed

Read each situation and choose an answer. The lesson explains the result. Try again if needed, then continue through all three decisions and finish the tour.

1. Choose a protected path

Your app needs to ask a model for a draft reply. The browser should not contain permanent AWS access keys. Choose where the model request should happen.

Which path keeps account credentials out of the page?

Browser → app server with a scoped role → Bedrock.

The server checks the user and uses a limited role to call the model. The browser receives the permitted result. This lesson is a design exercise and makes no paid model calls.

The app has a protected place to control model requests.

User → authorized server → permitted model request

2. Send only what is needed

A draft reply needs a short product description. A large file also contains customer passwords and private records. Choose what to send.

What input fits this task?

Only the necessary product details after privacy checks.

Minimize the data sent to a model and follow the applicable service settings and privacy rules. Do not send passwords or unrelated private data just because it is convenient.

The prompt contains the information needed for the draft.

Task → minimum suitable context → model prompt

3. Check before acting

The model drafts a refund reply, but the amount may be wrong. Its output should not automatically authorize a refund. Choose the next step.

Who should decide whether the action is allowed?

The app’s permission rules and a person where approval is required.

Treat model output as a draft, not a trusted command. Check facts and enforce permissions outside the model. Guardrails can help, but they do not guarantee correct answers.

The user reviews the draft before an important action is taken.

Model draft → fact check + authorization → approved action

Practice, then check your understanding

In the interactive tour, make each decision before seeing the outcome. Wrong answers receive coaching, and replay gives you a fresh attempt without duplicating earned XP.

Read the official AWS documentation for service behavior and production details.