How I Feel Pair Programming with AI

I believed AI was good for humanity, but when ChatGPT first launched I did not feel threatened. The code it generated rarely matched my taste, it often suggested obsolete methods that did not run, and I could not find a Neovim integration I liked.

So I skipped it for real work. I kept an AI code companion around for boring tasks like translating XML to TypeScript types or scaffolding a table, where it did a decent job and stayed out of my way.

Why I avoided AI code companions at first #

My taste matters when I read code later, and early ChatGPT rarely matched it. It also suggested APIs that no longer worked, so I spent time correcting the output instead of writing.

The bigger blocker was editor integration. I live in Neovim, and back then I could not find a companion that supported me without interfering while I typed.

What changed with ChatGPT-4o and a local runner #

After OpenAI made ChatGPT-4o available to VS Code users, I gave it another try and noticed a shift. The suggestions followed my style much more closely, and for popular languages like PHP, Python, and JavaScript they land correctly about 80% of the time. For newer or niche languages like Rust, I still see weaker suggestions and I double-check them.

That nudge pushed me to set up a local runner. I use Ollama with qwen2.5-coder:14b, and for Neovim I use codecompanion. The local setup responds quickly, it does not interrupt my flow, and I keep control over what gets sent out. This is the part that finally made pairing feel natural to me.

How I use AI now without letting it drive #

I let AI handle the parts where it saves me time and I keep the decisions. Here is how I use it most:

1. Generate tests that need mock setup. I ask AI to draft the test and the expectations when the test needs database or mock wiring. I review the assertions, I adjust the boundaries, and I keep responsibility for what the test proves. I admit I was slow to write these tests because the setup felt tedious, and AI removes that drag.

2. Learn a new library or language with your own code as context. I paste the code I am reading and ask how the library actually behaves. I have done this with Swift, PHP, Python, and Rust. The explanation is helpful, though with Rust I still verify the suggestion against docs before I trust it.

3. Handle repetitive transforms and small refactors. I use it to convert JSON to TypeScript types or to collapse a long else if chain. It gets the repetitive mapping right in most cases, and I do a final pass to name things the way I want them.

I do not use an AI editor like Trae yet, but the pattern feels similar. With this approach I ship about 50% faster because I no longer hand-write every test factory, so I spend that time on the actual feature.

The check that keeps me ahead of the model #

AI does not replace context. I read every generated block, I refactor it, and I keep the domain knowledge that the model lacks. If you try this, treat the output as a draft from a pair partner who types fast but does not know your system.

Try one of the three uses above on a small task this week, keep what reads like you wrote it, and send the rest back for revision. That habit has kept me productive without handing over judgment.