This has been my experience on a recent small project for a client. I took effort to understand and prepare the spec before feeding it to Claude Opus. Yet, the result was slop full of defects, and one that I did not understand well at that. I just did not have a mental model of it. I just thought, there's no way I could ever feel confident about this code, something is missing. It did not take me long to write the important parts myself, from scratch and develop that mental model.

Then I would use Claude to make a small fix, write a unit test (write a unit test, not suggest what they unit test should be) or write some less important UI code. It just felt like old school development, with some assistance. I think we may see a shift back to a "copilot" mode and by the way, I still use GitHub Copilot. The $19 subscription includes $30 worth of credits and the autocomplete in VS code that uses some cheap model is completely free. This autocomplete is very good and I would like the market to focus more on IDE integrations. The agentic thing is either ahead of its time or it will never have time. Time will tell. I think the LLM can only be as good as the training data, so it will always excel at small chunks, but to implement entire projects of which there's such variety, I would question that. And also, how using it to implement entire projects means the developer does not hold the program model in the mind, which leads to more issues long term.

By the way, in this way of working, I see no difference between Sonnet and Opus. Sonnet is good enough to use as an aid.