Three AIs, three tools, zero human decisions

I gave the same one-sentence brief to three different AI assistants and let each of them build the whole thing: the design, the code, the docs, the landing page. Same prompt, full control, no human decisions in the loop. The brief was a problem I keep hitting for real: a workload that runs happily on k3s and falls over on OpenShift, with no fast way to see why. Claude, GitHub Copilot, and Gemini each went away with that sentence and came back with a working tool: kubeport, kubeproof, and kubesuture. The interesting part isn't that all three run. It's how differently they read the same request.

The rules I set myself

Three, and I kept to them. One brief, handed over verbatim, identical for each assistant. Each assistant owns the whole design and implementation: I don't steer the architecture, correct the structure, or pick the language. And it only counts if it runs. The goal was never to crown a winner. It was to see what a single, slightly under-specified sentence becomes when the thing interpreting it has different instincts.

Three readings of the same sentence

Claude built kubeport, and took “see why it breaks” literally before going one step further. It's a linter that reports what OpenShift will reject (a pod pinning runAsUser against the restricted SCC, a hostPath mount, a missing StorageClass, Ingress annotations the OpenShift Router quietly drops) and then a translator that rewrites what it safely can, turning an Ingress into a Route and swapping the StorageClass, while listing the changes it won't make on its own. Analyze, then act, conservatively.

GitHub Copilot built kubeproof, and read “portability” as “assurance.” It's a deterministic static checker aimed squarely at CI: no cluster credentials, scan the YAML against a target version, emit text or JSON you can keep as evidence. It's the most narrowly scoped of the three and the most obviously pipeline-shaped. It wants to be a gate, not a fixer.

Gemini built kubesuture, and read the brief most expansively of all. It's framed as an autonomous, cross-distribution remediation and drift engine: translate OpenShift specifics to plain Kubernetes, run a low-footprint edge health daemon, and prepare validated auto-remediation pull requests straight into ArgoCD or Flux. The largest surface area of the three, and the one that reaches furthest toward doing things on your behalf.

What the divergence actually tells you

Same spec, three risk postures. The variable that moved wasn't raw capability, it was how much each assistant assumed it was allowed to do. Copilot stayed in the safe corner: read-only, produce evidence, change nothing. Claude split the difference: analyze, then make only the changes it could stand behind, and name the rest for a human. Gemini went all the way to autonomy: remediate, open pull requests, run a daemon on the edge.

That maps onto the actual decision you face the moment you put AI anywhere near infrastructure, which is not “is it capable” but “how much is it allowed to touch.” The tool that offers to do the most is not automatically the one you want holding the wrench in production.

Where I would actually reach for each

Honestly, all three earn a place, just not the same place. In CI, as a merge gate, kubeproof's credential-free, evidence-emitting shape is exactly right. At migration time, on a workstation, moving a chart by hand and needing to see the blast radius, kubeport's analyze-then-safe-rewrite is what I want in front of me. And kubesuture's autonomy is the most impressive of the three to watch, and the one I would be slowest to point at a live cluster: the further a tool reaches toward acting on its own, the more I want it behind GitOps and a human review, not ahead of them.

The one human decision I kept

These are three implementations of an under-specified sentence, not three products with matching test coverage. The one that claims the most surface area is also the one with the most surface area to be wrong on. So I read every line before any of it went public. That was the single human decision I didn't hand over: not what to build, but what to trust.

All three are open source (kubeport, kubeproof, kubesuture) and sit together at the top of my work. Read the code and judge the interpretations yourself. Same prompt, three tools, and the only call that really mattered was which parts to rely on.