I have been trying LLMs in a couple of new small projects recently. I got more success that I hoped for, but I had to adjust my usage to be effective. First o

I have been trying LLMs in a couple of new small projects recently. I got more success that I hoped for, but I had to adjust my usage to be effective. First of all, treat the LLM as a less experienced programmer. Don't trust it blindly but always make code review of its changes. This gives several benefits. 1) It keeps you in touch with the code base, so when need arise you can delve into it without too much trouble 2) You catch errors (sometimes huge ones) right away, and you can have them fixed easily 3) You can catch errors on your specification right away. Sometimes I forget some detail and I realize it only when reviewing, or maybe the LLMs did actually handle it, and I can just tell it to update the documentation 4) You can adjust little by little the guidelines for the LLM, so that it won't repeat the same "mistakes" (wrong technical decisions) again. In time you get a feeling of what it can and cannot do, where you need to be specific and where you know it will get it right, or where you don't need to go into detail. The time required will be higher than vibe coding, but decreases over time and still better than doing by myself. There is another important benefit for me in using an LLM. I don't only write code, I do in fact many other things. Calls, writing documentation, discussing requirements etc. Going back to writing code requires a change of mental state and to recall into memory all the required knowledge (like how is the project structured, how to use some apis etc.). If I can do two hours of coding it is ok, but if the change is small, it becomes the part where I spend the majority of time and mental energy. Or I can ask the LLM to make the changes and review them. Seeing the code already done requires less energy and will help me reminding stuff.