Our team is using Claude Code on Terraform for AWS, but changes still miss acceptance criteria. Reviews with different AI tools keep turning up more issues, even after fixes. How do you set a useful team workflow and know when a change is ready to merge?
How do you make AI-assisted DevOps work as a team without endless reviews?
Answers
Make as much of the acceptance criteria executable as you can. For Terraform, that might mean formatting and validation checks, TFLint, policy checks against the plan with OPA or Conftest, and integration tests for the behavior you actually expect.
Run those checks locally or in CI before asking someone to review. A green pipeline is not proof that infrastructure is safe, but it gives the team a repeatable baseline. Keep a human review for the plan, the risks, and anything the tests cannot establish.
Reply to this answer
Stop treating zero AI findings as the goal. A fresh review can generate another plausible concern even when the code hasn't changed, and the severity label is not a measurement by itself.
Set a review bar before starting: acceptance criteria, security and operational risks, and anything your team considers a blocker. Ask the reviewer to explain each finding and point to the relevant code. A person can then fix, reject, or defer it. After a pass and any necessary fixes, review the actual changes rather than starting an unlimited cycle of broad reviews.
Reply to this answer
The person submitting the change still owns it, whether they typed it or a tool generated it. They should be able to explain what the Terraform plan will do and why the change meets the ticket. If they can't, that needs resolving before merge, not another round of agents talking to each other.
Reply to this answer
For team consistency, put the recurring instructions and checks somewhere everyone can use them. Document the relevant infrastructure conventions, common workflows, and commands alongside the scripts that enforce them. Feed that context to the tool for the task at hand instead of relying on each person to recreate the same prompt.
Keep the guidance under review too. A shared instruction can spread a bad assumption just as efficiently as a good one.
Reply to this answer
You don't have to give the tool deployment access to get value from it. Try it for explaining a codebase, investigating problems with read-only access, drafting scripts, or suggesting changes in a restricted workspace. A human can inspect the result and run the plan or apply separately.
_f37!parcelruler.8:
Even read-only access can expose sensitive configuration or data, so the scope of that access still needs thought.
Reply to this answer
Add your answer
Answers are reviewed before publication.
_m84+gaugeparcel.6:
It also helps to tell the coding tool which checks to run and what output would demonstrate each criterion, rather than just giving it a prose checklist.