← Tatia Razmadze

The time it was wrongand the time it wasn’t

I write tooling that reviews infrastructure pull requests. It advises; I approve. Here is one of each.

It was wrong

A wildcard that had to be a wildcard

It flagged Resource = "*" in an IAM policy as overbroad. Usually that’s a fair catch — it’s the most common finding any linting tool makes.

Not here. That AWS action doesn’t support resource-level permissions at all, so "*" is the only value it accepts. Narrowing it wouldn’t have made the policy safer. It would have made it invalid.

I checked the service authorization reference rather than arguing from memory, then looked at where the boundary actually lived — the OIDC trust conditions on the role. Already tight.

So I dismissed it and wrote down why. Not “false positive” — the real reason, in the pull request, so the next person doesn’t redo the research.

It was right

A database nobody said to keep private

A change created an RDS instance in valid, working Terraform that never set publicly_accessible. Nothing was syntactically wrong. The requirement simply wasn’t expressed anywhere in the code.

A human review would probably have passed it — you skim a diff, you don’t re-read the acceptance criteria line by line.

I hit that gap, so I encoded it. My rds-terraform-provisioner skill now carries it as a gotcha: requirements go quiet on that setting, so set it explicitly rather than inheriting whatever the default happens to be.

Reviewing work you didn’t write is an ordinary engineering skill. Agentic tooling just makes it load-bearing.

The interesting question in an interview isn’t whether I use these tools. It’s whether I can tell you about a time one of them was wrong, and what it cost to find out.