That said, everytime I give it something a little more complex that do something in a single file script it fails me horribly. Either the code is really bad, or the approach is as bad a someone who doesn't really know what to do or it plains start doing things that I explicitly said not to do in the initial prompt.
I have sometimes asked my LLM fan's coworkers to come and help when that happens and they also are not able to "fix it", but somehow I am the one doing it wrong due "wrong prompt" or "lack of correct context".
I have created a lot of "Agents.md" files, drop files into the context window... Nothing.
When I need to do green field stuff, or PoCs it delivers fast, but then applying it to work inside an existent big application fails.
The only place where I feel as "productive" as I heard from other people is when I do stuff in languages or technologies I don't know at all, but then again, I also don't know if that functional code I get at the end is broken in things I am not aware of.
Are any of you guys really using LLMs to create full features in big enterprise apps?
linesofcode•4h ago
What is health of your enterprise code base? If it’s anything like ones I’ve experienced it’s a legacy mess then it’s absolutely understandable that an LLMs output is subpar when taking on larger tasks.
Also depends on the models and plan you’re on. There is a significant increase in quality when comparing Cursors default model on a free plan vs Opus 4.5 on a maximum Claude plan.
I think a good exercise is to prohibit yourself from writing any code manually and force yourself to do LLM only, might sound silly but it will develop that skill-set.
Try Claude code in thinking mode with the some super powers - https://github.com/obra/superpowers
I routinely make an implementation plan with Claude and then step away for 15 mins while it spins - the results aren’t perfect but fixing that remaining 10% is better than writing 100% of it myself.
not_that_d•3h ago
Given it to new people of course carry questions, but most of them (juniors) could just follow the code given an entry point for that task, this from BE to FE.
I use the github copilot premium models available.
> I routinely make an implementation plan with Claude and then step away for 15 mins while it spins - the results aren’t perfect but fixing that remaining 10% is better than writing 100% of it myself.
I have to be honest, I just did this two times and the amount of code that needed to be fixed, and the mental overload to find open bugs was much more than just guide the LLM on every step. But this was a couple of months ago.
not_that_d•3h ago