Comment by xyzzy123
1 year ago
For easy things, LLM assist has sped things up a lot for me.
For medium complexity things, I can get them done quickly without manual coding if I have a clear understanding in mind of what the implementation should look like. I supply the requirements, design and strategy and it's fairly easy to "keep things on the rails". The "write a PRD first" hack (https://www.aiagentshub.net/blog/how-to-10x-your-development...) works pretty well. Agent with YOLO mode and terminal access rips, particularly if you have good tests.
For tasks where I know the spec of the feature but don't clearly understand how I would design / implement the feature myself, it's hit-and-miss. Mostly miss for me.
I also haven't had much success with niche libraries, have to stick to the most popular library/tool/framework choices or it will quickly get stuck.
Whereas I've been disabling AI assist features because I find them actively disruptive to the development process. When it ghost pops up text suggesting what I should do, it's sometimes right...but it breaks flow. It forces me to read and parse apparently correct code, and decide if it is correct or it's just a mirage which is valid but not actually what I'm doing at all.