A look at how our engineering team uses AI coding tools while building mirrord, including what works, what doesn't, and real-world examples.

MetalBear

MetalBear's engineering team shares practical experiences using AI coding tools while building mirrord, a Kubernetes development tool in Rust. AI proves most valuable for understanding unfamiliar code, exploring architectural alternatives, and generating scripts. It struggles with complex architectures and long-running reasoning tasks. The team finds ChatGPT most reliable for iteration, Gemini best for deep research but prone to losing context, and Claude Code somewhere in between. Success depends on scoping problems tightly and controlling context rather than which model is used. AI accelerates tedious and exploratory work but cannot replace deep system understanding.

How Our Engineering Team Uses AI

So is AI changing how we build software at MetalBear? #

I heard people saying that AI is exceptionally well at frontend/backend coding, but subpar at infrastructure config making. So at that area its not a surprise to see such a low bar.

I have been hearing about coding with AI for some time now, but no one is talking about how to implement and develop tests for that code with AI.

How would you think AI would behave if you constantly shared conversations? E.g. in each commit description a conversation between user and AI is given. When, before doing changes, AI could look at git history, find relevant commits and read context. Thst could greatly reduce the chances that AI would break domething due to lack of context.
Same applies if you have something like Confluence: if AI can query the docs and they are well structured, AI efficiency increases.
I believe the fault is in humans not providing/reminding AI right context at the right time, expecting AI to guess requirements out of thin air or from the mass codebase with little support.
As AI can code quickly we should gocus on effectively summarizing/documenting data and keeping docs constantly up to date.