How to use Claude to investigate failed AWS deploys
In the reported incident, the investigator flagged a missing Fargate environment variable; the developer added it and redeployed without asking the DevOps engineer.
dev.to
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Investigate failed CloudFormation stack [stack name] with optional suspected commit [commit]. Use [read-only AWS MCP role] to read stack state. Produce a short analysis with evidence and ranked hypotheses if unsure, then send it to [configured output destination].
Reach people exploring AI agents
See placements →What happened
In the reported incident, the investigator flagged a missing Fargate environment variable; the developer added it and redeployed without asking the DevOps engineer.
Reported by deeheber on dev.toHow the agent did it
- Built a CodeBuild project that ran Claude Code headlessly.
- Passed a failing stack name to inspect state through read-only AWS MCP.
- Generated a short analysis and posted it into the alert’s Slack thread.
Where a person stepped in
A developer read the diagnosis, added the missing environment variable, and redeployed. The source notes broad read-only IAM permissions.
A person set the goal and reviewed the result.
Danielle Heberling built cfn-investigator so teammates could get a starting point for failed CloudFormation deployments without waiting for the team's sole DevOps engineer. The tool ran Claude Code headlessly in AWS CodeBuild and accepted a failing stack name, optionally accompanied by a suspected commit.
Claude read stack state through an AWS MCP server using a read-only role, worked out likely causes, and wrote a short analysis. The system prompt allowed an unsure response with ranked hypotheses rather than forcing a definitive answer. A public example logged results to CloudWatch and provided a forwarding point; the author's workplace version posted into the Slack thread containing the deployment alert.
The author used CodeBuild and an Anthropic API key to reach a working prototype quickly. They acknowledged broad ReadOnlyAccess permissions, tools installed anew each run, and a role split that scoped MCP AWS calls rather than Claude itself.
In one reported incident, the investigator flagged a missing environment variable on a Fargate task definition. The developer read the message, added the variable, and redeployed without contacting anyone for help. The tool supplied diagnosis rather than repairing the deployment.
Who it's for: Teams wanting an automated first investigation of failed stacks while keeping deployment changes with developers.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: dev.to · deeheber
Frequently asked questions
Does the investigator repair deployments?
The author says it provides what broke and where to start. In the reported case, a developer added the missing variable and redeployed.
How does it handle uncertainty?
Its system prompt includes an unsure option that ranks hypotheses instead of inventing a clean answer.
More details
Remaining steps
- Used the reported missing variable to correct and redeploy a Fargate task.
About this account
Claude investigated stack state and produced analysis automatically; it did not fix the deployment.
Platform: dev.to. Attribution: https://dev.to/aws-heroes/getting-claude-code-off-my-laptop-and-onto-shared-compute-4cjc. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: deeheber on dev.to
Source published 23 May 2026
Source checked 14 Oct 2026
Tools and topics
Claude Code · AWS CodeBuild · AWS MCP server · CloudWatch · Slack
CloudFormation · Deployment diagnosis · Headless agents
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.