There's nothing "hollowed out" about directing an AI effectively, the feedback is as quick and tight as it always was. The trick is that you don't just "vibe code" and let the AI one-shot the whole thing: you should propose the change first and ask the AI about a good, detailed plan for implementing it. Then you review what the robot has proposed (which is trivial compared to revising code!) make sensible changes, ask for feedback again, and repeat. By the time the AI bot has to write actual code, it's not running on vibes anymore: it's been told exactly what to do and how to assess the result. You spend more time upfront, but a lot less on fixing the AI's mistakes.

▲

skydhash 2 hours ago | parent [-]

> you should propose the change first and ask the AI about a good, detailed plan for implementing

Why ask though?

If I’m familiar with a project, more often than not, I usually have a very good idea of the code I have to write within minutes of reading the ticket. Most of the time taken is finding the impact of the change, especially with dependencies that are present in the business domain, but are not reflected in the code.

I don’t need to ask what to code. I can deduce it as easily as doing 2+2. What I’m seeking is a reason not to write it the way I envisioned it. And if those reasons are technical, it’s not often a matter of code.

▲

zozbot234 2 hours ago | parent [-]

Because that's how you ensure that the AI has the right idea about what to do. If the proposed plan has problems, you work with the AI to fix them before setting it to work. AI is not as smart as you, so it needs to be told how to go about doing things.

▲

skydhash 2 hours ago | parent [-]

Any change that I’ve done which resulted in more than a a 10 lines diff are done with tools (copy-paste, vim-fu, refactor tools or script, snippets, code generators,…) Why would I spend time babysitting an LLM when I could have just done it myself? The purpose of automation is to lighten my workload, not to add to it.

▲

zozbot234 an hour ago | parent [-]

An LLM is also a code generator. There is a scale of changes where using one is just not worthwhile (quite possibly around the 10 lines mark, as you said) but other than that, why would you want to write code yourself line-by-line that you could just generate?

	▲	skydhash 39 minutes ago \| parent [-]
		Who even write their code line by line? Snippets and other code generation tool has been here for decades. If you’re writing Java in IDEA, it’s basically a tab-fest with completion. And if you’re fluent in your editor, you do much more complex than editing lines.