Bedrock Prompt Management: Versioned System Prompts for Production

System prompts buried in Lambda env vars are unreviewable and unrollbackable. Treat them like code: create versions in Bedrock Prompt Management (or your own registry), attach aliases (dev, canary, prod), evaluate quality/latency, and roll back in one alias update when assistants regress.

⚡ TL;DR: Store prompts as versioned artifacts; never hot-edit prod strings in the console without a version bump; canary an alias on 5–10% of traffic; compare eval scores and p95 latency; flip alias back on regression. Pair with prompt caching, RAG eval (sibling angle), and Bedrock Agents.

Create and version prompts as code

import boto3
bedrock = boto3.client("bedrock-agent")

def publish_prompt(name: str, text: str, desc: str) -> dict:
    # First create, then create version — illustrative API shape
    created = bedrock.create_prompt(
        name=name,
        description=desc,
        variants=[{
            "name": "default",
            "templateConfiguration": {
                "text": {
                    "text": text,
                    "inputVariables": [{"name": "service"}, {"name": "locale"}],
                }
            },
            "modelId": "anthropic.claude-sonnet-4-20250514-v1:0",
            "inferenceConfiguration": {"text": {"temperature": 0, "maxTokens": 2048}},
        }],
    )
    versioned = bedrock.create_prompt_version(
        promptIdentifier=created["id"],
        description=desc,
    )
    return versioned

Keep the source of truth in git (prompts/oncall-v3.txt) and publish from CI — console edits are break-glass only.

Aliases for canary and prod

def point_alias(prompt_id: str, alias: str, version: str):
    # create_prompt_alias / update_prompt_alias patterns
    return bedrock.create_prompt_alias(
        promptIdentifier=prompt_id,
        promptVersion=version,
        name=alias,  # prod | canary | dev
        description=f"points to {version}",
    )

// Runtime resolves alias, not hard-coded version
const promptArn = process.env.PROMPT_ALIAS_ARN!; // ...:prompt/xxx:alias/prod
// Pass prompt ARN into Converse / InvokeModelWithResponseStream as supported

✅ Alias flip = instant rollback.
❌ Redeploying twelve services to change a paragraph.

Canary with eval gates

# .github/workflows/prompt-canary.yml
on:
  workflow_dispatch:
    inputs:
      version: { required: true }
jobs:
  eval:
    runs-on: ubuntu-latest
    steps:
      - run: python eval/run_golden.py --prompt-version ${{ inputs.version }} --out out.json
      - run: python eval/compare.py --baseline s3://…/baseline.json --candidate out.json --min-faithfulness 4.0 --max-p95-ms 3500
  promote:
    needs: eval
    steps:
      - run: python prompts/point_alias.py --alias canary --version ${{ inputs.version }}
      - run: python prompts/bake.py --alias canary --minutes 60
      - run: python prompts/point_alias.py --alias prod --version ${{ inputs.version }}

Use the same golden-set discipline as RAG evaluation. Cache stable prompt prefixes — Bedrock prompt caching.

Observability

Emit prompt_alias, prompt_version, latency, tokens, user thumbs-down. Alert when thumbs-down rate or tool-error rate spikes after an alias change.

Closing checklist

✅ Dos
– ✅ Git-source prompts; publish versions from CI
– ✅ Use aliases (dev/canary/prod), never hard-code versions in ten services
– ✅ Gate promotion on golden eval + latency
– ✅ One-command alias rollback
– ✅ Log prompt version on every request

❌ Don’ts
– ❌ Don’t edit the live prod prompt text in-console without a new version
– ❌ Don’t ship prompt changes bundled silently with unrelated app deploys
– ❌ Don’t skip canary on high-traffic assistants
– ❌ Don’t raise temperature when versioning “for creativity” in prod support bots
– ❌ Don’t forget input variables — broken templates fail open into empty system prompts

Related reading

• Bedrock Prompt Caching and Batch Inference: Cut Latency and Cost
• Amazon Bedrock Agents: Tool Use, Memory, and Production Guardrails
• RAG on AWS: OpenSearch vs Aurora pgvector for Codebase Chat
• AI Code Review Bots: IAM, Secrets, and Least-Privilege Pipelines

Last updated on September 11, 2026