←back to thread

86 points shandsaker_au | 1 comments | | HN request time: 0.199s | source
Show context
patapong ◴[] No.45166528[source]
> The products I built with Claude are worse than without them because I use programming as a way to think and interact with the problem. When you're coding, you're deeply invested in the problem you're solving, getting intimate with the problem. With AI tools, it's surface level. It's a one-night stand with a problem versus a deep and meaningful relationship.

A very interesting insight about AI coding. It gets at the theory building part of programming, which is much harder to do when just supervising an AI in my experience. On the other hand, I am so much faster that it's hard not to use AI for coding. Interested to see what they come up with!

replies(1): >>45167538 #
1. kannanvijayan ◴[] No.45167538[source]
I'd been struggling to find good use cases for agentic things like Claude for this reason. My project is not a good fit for it as there's enough novelty to throw off the ML pattern systems. I was benefitting greatly from autocomplete, but not really leaning too much on code-writing tools.

But I actually found a use case where the agentic approach adds incredible value: internal tooling and visualization.

I was debugging some feature in the core, and had some code to jump a JSON diagnostic structure with a lot of info. Looking at this directly was getting to be a pain, so I wanted to write a UI for it.

Claude handled this task almost perfectly for <$5. I wrote up a description of the JSON schema, and a description of how that should map to a view, added some styling directions, and let er rip. In one afternoon, I go the UI built for me when I was doing dishes.

So I built the tool and went back to debugging and made a ton of progress using it. It's ok if I don't have a full internalization of the visualization architecture. It's basically a complex software "jig" that the AI built for me. I can build new ones as necessary.

A very interesting revelation.