gfwiwcgws

I once spent two hours rewriting an if-else loop hell into a binary search tree.

This was back in the day, before LLM-assisted coding could do the heavy lifting for you. I was building a decision rules engine. The if-else chain was a raft taking on water, and the BST was the textbook, computer-science-approved answer. Two hours of manual typing. Of course, the BST passed every test. It was "good" code.

But "good" is a stupid prompt.

Code isn't just good or bad in a vacuum. It is gfwiwcgws. If you don't name the context, you're just arguing with ghosts. You're chasing whales when you should be patching the raft.

If you want to know if a tool or a refactor is actually the right move, write a failing test for the standard. Don't measure elegance. Measure the boundary. Input and output. When an input is provided, does the chosen tool produce the exact same result every single time? That is the assertion. The test is the lever. It pins the standard down so the codebase has to honor it, every commit, forever.

This is how you stop the religious wars. The ones about whether to rent intelligence by the token or run it on local metal. You stop arguing about what is universally "best" and start looking at the return on investment for your specific context. You invest early in better, cheaper storage. You put the model on your own disk, next to your dev machine, because you know the ROI is going to be undeniable. It will be worth it. Just wait and see.

Name the context. Write the test. Watch it go red.