1. App Frontend Development Since AI can't see screens yet, I have the advantage of having eyes and need to do this myself. The app development environment is much more unstable than web development, so it takes quite some time to attach something like MCP to it. I often feel discouraged that I'm doing this myself, especially since AI probably has better design sense than I do (not because AI is a superhuman designer, but because I'm very subhuman in this regard).

2. Backend Development Thanks to libraries like Pothos and Prisma, I can quickly handle simple CRUD operations myself. Of course, AI would do it faster, but most of this is already developed. What I really want to delegate is developing complex and tricky mutations, but there's a problem: I haven't set up a testing environment in advance. The code AI produces often looks correct at first glance, but upon closer inspection, it frequently fails to meet specifications. I need to quickly create e2e tests and incorporate AI into the feedback loop. With some effort in this area, I could receive much more help from AI.

3. Library Development There are parts of my current frontend/backend code where introducing libraries that generalize problems would greatly improve code quality. I have a direction in mind for how to create these. So while I was directly developing the product with my own hands, I tried having Windsurf create libraries in agent mode, but the results weren't good. I gave it design instructions, had it write test code, and told it to fix things until the tests passed, kicking it whenever it broke down and stopped, but I had to kick it so often that I got tired and gave up.

However, automating library development carries a dangerous implication. Libraries are "accumulating" code. They're different from product code that you hastily pile up and ship before it topples sideways. If AI could develop libraries on its own, there would truly be nothing left for humans to do. Ironically, this is why tools like Cursor and Windsurf are selling for billions - because AI still can't develop libraries, so individuals pay for hundreds of thousands of tokens to write product code that will eventually topple sideways. A recent post on Hackers' Pub also suggests that instead of wasting tokens on short-term goals individually, we should combine forces to accumulate code. In reality, if this works well, it would pose a significant long-term threat to coding agent businesses (and many other things).

But to develop libraries, we need Zen AI that surpasses Gen AI. Whatever, someone will create it soon enough.