Ask HN: Is anyone experimenting with different ways of using LLMs for coding?

One fed-up coder asked a very relatable question on Hacker News: why does coding with artificial intelligence still feel so awkward? Instead of gliding through work, they said using tools like Claude and Codex feels like riding a bike that slams the brakes every few minutes. That image absolutely set the mood, and the crowd pounced. The big feeling in the room was clear: people are intrigued by AI for programming, but a lot of them are also deeply annoyed by the stop-start routine of typing instructions, waiting, checking mistakes, and starting over.

The hottest divide? One camp basically said, stop trying to build elaborate hacks around the AI and just write better instructions. In their view, the model is already smart enough; humans are the messy part. Another camp said, not so fast — they’re building custom systems precisely because the current chat loop is exhausting. One commenter is making a template language just to keep big projects from turning into context spaghetti, while another confessed they now keep a giant free-text idea file and periodically tell Claude, in essence, “hey, I updated the list,” which is both ingenious and hilariously low-tech.

Then came the dreamers chasing a smoother future: scratchpads, better autocomplete, and tools meant to remove those infamous “brakes.” The accidental comedy of the thread is that everyone wants AI to feel magical, yet half the comments read like people inventing sticky notes, clipboards, and better steering wheels for a bicycle that still won’t coast.