[htx80nerd]

>The cost of producing code has collapsed. AI tools can generate functional, adequate, perfectly average code at a speed and cost that would have been unimaginable even five years ago. And like the outsourcing wave of the early 2000s, the economics are real and rational. Nobody is wrong for using these tools. The code they produce is often fine. It works. It passes tests. It might ship as-is.

After using AI for months (Claude, Gemini, ChatGPT) it is extremely rare for their code to work 'as is' first shot and almost always requires several iterations and cleaning up edge-cases.

When it does work 'first shot' it's usually when it's transferring existing working code to a new project which is slightly different.

[simonw]

Have you tried the "use red/green TDD" trick?

I believe that increases the chances of one-shot code working, though it's also possible that it did that against Opus 4.5 and isn't necessary against Opus 4.7 but I haven't spotted the difference yet.

[sanderjd]

Yeah this is the way. Thinking about how it will verify that it has done the right thing is the key. This been be set up at the AGENTS.md level.

Very simple things like: "Write tests and make sure they pass." "Run lint after each change." "Write API docs in XYZ format."

In my experience, they are very good at fixing things they've done wrong after discovering them during those kinds of steps.

[sanderjd]

Weird, this isn't my experience at all (mostly writing Python lately). Granted, it usually doesn't implement things exactly the way I want them to be implemented, and I iterate a lot on that. But I think it's been like a year? at least six months, since the code didn't work on the first try.

[bluebands]

try gpt-5.5-xhigh fast in codex mac app, preferably with TDD and /goal, with a clearly defined end result

it's unbelievable, it will do the iterations for you, it will easily work 12 hours straight until it's a good output