My team has started using Claude Code and other AI coding tools to work on infrastructure as code for AWS. Our stack includes Terraform, EKS, Aurora, GitOps, Argo CD, CloudFront, API Gateway, and S3. We write tasks with detailed acceptance criteria, but the generated code still sometimes misses them. When I ask Claude to review a merge request, then use expert review or Codex to check the same changes, each tool finds additional issues. After merging the feedback and asking Claude to fix everything, another review often discovers more problems. The process feels endless. How are other teams using AI effectively while keeping reviews focused and maintaining code quality?
4 Answers
Make the acceptance criteria machine-checkable wherever possible. Have the team agree on the specification and review that before implementation, then give the agent access to reliable Terraform and AWS documentation rather than relying on its memory. Run formatting, validation, security checks, Terraform plans, tests, and a targeted AI review in CI before spending human review time. A prompt like “find every possible issue” tends to produce an endless stream of low-value suggestions, so review against a defined checklist instead.
Treat the AI like another tool in the developer’s toolbox, not as the person responsible for the result. The engineer who submits the change still needs to understand the Terraform and AWS behavior, explain the design, and own the code. “The model generated it and I don’t know how it works” shouldn’t be an acceptable explanation except for truly trivial scripts.
A ticket-driven workflow can help: put the design, assumptions, acceptance criteria, and review checklist in the ticket first. Let the team challenge that plan before asking an agent to implement it. Then the agent works on a bounded task, and the merge request is evaluated against the original ticket instead of being repeatedly reopened based on whatever a new model happens to notice.
The endless loop usually happens because no one has defined what “good enough to ship” means. Set a fixed bar: the acceptance criteria pass, automated checks are clean, required security and reliability concerns are addressed, and a human approves the change. Use one or two focused review passes, not a chain of agents reviewing and rewriting each other. Once the agreed checklist passes, make a deliberate ship-or-no-ship decision and stop chasing stylistic nitpicks.
This applies to human-written code too. AI reviewers can still find something to criticize in any codebase, so the goal should be risk reduction against explicit requirements, not proving that the code has zero possible flaws.