Sure. Another article was posted today[1] on the subject. An example claim:
> If we asked the AI to solve a task that was already partially solved, it would just replicate code all over the project. We’d end up with three different card components. Yes, this is where reviews are important, but it’s very tiring to tell the AI for the nth time that we already have a Text component with defined sizes and colors. Adding this information to the guidelines didn’t work BTW.
This is helpful framing. I would say to this: I have also noticed this pattern. I have seen two approaches help. One, I break up UI / backend tasks. At the end of UI tasks, and sometimes before I even look at the code, I say: "Have you reviewed your code against the existing components library <link to doc>?" and sometimes "Have you reviewed the written code compared to existing patterns and can you identify opportunities for abstraction?" (I use plan mode for the latter, and review what it says). The other approach which I have seen others try, but have not myself (but it makes sense), is to automatically do this with a sub agent or hook. At a high level it seems like a good approach given I am manually doing the same thing now.
> I am still surprised at things it cannot do, for example Claude code could not seem to stitch together three screens in an iOS app using the latest SwiftUI (I am not an iOS dev).
You made a critical comment yet didn't follow your own rules lol.
> it's so helpful for meaningful conversation!
How so?
FWIW - I too have used LLMs for both coding and personal prompting. I think the general conclusion is that it when it works, it works well but when it fails it can fail miserably and be disastrous. I've come to conclusion because I read people complaining here and through my own experience.
Here's the problem:
- It's not valuable for me to print out my whole prompt sequence (and context for that matter) in a message board. The effort is boundless and the return is minimal.
- LLMs should just work(TM). The fact that they can fail so spectacularly is a glaring issue. These aren't just bugs, they are foundational because LLMs by their nature are probabilistic and not deterministic. Which means providing specific defect criteria has limited value.