What's changed: Initial: 5 sections for Domain 1 (FM integration, data management, compliance)
1.5Prompt Engineering Strategies and Governance
Learn prompt strategies that control FM behavior and the governance around them: Bedrock Prompt Management (templates/versions/variant comparison), Bedrock Guardrails, chain-of-thought, Bedrock Prompt Flows, and auditing via CloudTrail/CloudWatch Logs.
Prompts are “code”—subject to versioning, testing, and approval. At the professional level, beyond effective prompt design you must apply organizational governance (who changed which prompt, and how).
1.5.1Prompt design and governance
- Bedrock Prompt Management: centralize prompts with parameterized templates, versioning, and variant comparison (gate production promotion with an external step, e.g., CodePipeline).
- Bedrock Guardrails: enforce Responsible AI rules (role definitions, denied topics, PII filters) at the prompt layer.
- Advanced techniques: chain-of-thought, structured inputs, output-format specs, and feedback loops to raise quality.
- Bedrock Prompt Flows: visually compose sequential prompt chains, conditional branching, reusable components, and pre/post-processing for complex tasks.
- Auditing: track usage with CloudTrail and log access/prompts with CloudWatch Logs.
Common: prompt templates/versions/variant comparison = Prompt Management, enforce Responsible AI rules = Guardrails, compose sequential chains/branching = Prompt Flows, improve accuracy with reasoning = chain-of-thought, audit prompt usage = CloudTrail/CloudWatch Logs.
The core is prompt lifecycle management. In Bedrock Prompt Management, build parameterized templates, save changes as versions, and promote to production via version → alias, gating with an external approval step (e.g., CodePipeline) when required. Automate prompt regression testing with Lambda (verify expected output), Step Functions (edge cases), and CloudWatch (monitor prompt regressions). Control behavior with Bedrock Guardrails (block role drift, denied topics, PII) and structured output (JSON Schema). Compose complex tasks in Bedrock Prompt Flows (prompt → KB → Lambda), promoting via working draft → immutable version → alias. Audit usage with CloudTrail (API calls) and CloudWatch Logs (prompt/response logs) to track who used which template when. Treat prompts not as disposable plaintext but as assets with templates + versions + variant comparison + auditing.
| Goal | Use | Point |
|---|---|---|
| Centralize prompts | Prompt Management | Templates/versions/variant comparison |
| Safe behavior control | Guardrails | Denied topics/PII/role |
| Compose complex tasks | Prompt Flows | Chains/branching/versions & aliases |
| Audit usage | CloudTrail/CloudWatch Logs | Who/when/what |
Trap: “hard-code prompts in the app and edit them ad hoc” is a governance failure—manage them with Prompt Management templates + versions + variant comparison. Also “Guardrails is a prompting technique that boosts accuracy” is wrong—it enforces safety; accuracy techniques are things like chain-of-thought.
1.5.2Section summary
- Prompts are assets = templates + versions + variant comparison (Prompt Management) + auditing
- Safety = Guardrails / composition = Prompt Flows / accuracy technique = chain-of-thought
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. You want to template and version production prompts used by multiple teams and compare variants to promote changes. Best option?
Q2. You must enforce safety rules at the prompt layer so a customer bot avoids denied topics and PII output. What?
Q3. You want to compose a multi-step “prompt → knowledge base lookup → Lambda post-processing” flow with branching and promote it as an immutable version. Best option?