Will start in a weird way – LLMs are very poor when it comes to basic logic and reasoning. Why? Because they were never designed to be good at logic. They were designed to excel in NLP (Natural Language Processing) and in that – they do excel. LLMs are fundamentally limited by their design and underlying drive – Mathematics. One can think of LLMs as products of linear algebra, probability and statistics, pinch of calculus and some neural processing network (hail the transformer). Stating all this because understanding their limitations allows one to work around them and use LLMs the proper way, the one which yells the desired result.
LLMs have become good at coding, the arbitrary date – mid 2025. When I say ‘good’ think about an average engineer with medium experience. The one that does not (and will not) excel but one that can still be relayed on most of the time. LLMs can do the heavy lifting, liberating one from the task of writing the code itself, while allowing for rapid implementation and deployment. Think of an LLM as a not very bright but sort of cooperative chef which can cook very well while guided. If all else equal (the LLM is good enough) the quality of the food would be (mostly) in direct relation to the provided instructions while the speed and accuracy of execution would be mostly thanks to the agent.
If you are a good engineer, who loves solving problems and knows his shit, then with a proper use of LLM you can, literally, amplify your skills and capabilities by offloading the actual work on the agent(s) while focusing on the design, code reviews and overall end result. The better you are – the better the end result would be. If you are a mediocre or sloppy, so would be the result. The same applies for virtually any engineering and business role. LLMs are a powerful skill amplifiers if used right. On the contrary – they will not help the lazy ones, the ones who are not familiar with the fundamental limitations, the usage, the paradigm, the foundational abstracts. Those would still suck big time because there is no cure for the voluntarily retarded.
It’s all about context
End of the day that’s all it is about – providing proper context in the form of a mega prompt that is broken down and described in various MD files – each targeting a specific area / domain that needs proper approach. To do this one must think architecturally, have overall good view over the “big picture”, be strategic, tactical and sometimes operational. A basic setup would be – using MCP (Model Context Protocol) to allow the agent (say Claude code) access to the relevant data (you must decide which one it is) and code base, do your MDs and start with something simple. Prepare for a lot of mistakes, prepare for a lot of resets and do-overs and incremental improvement. If you know what to accomplish, if you are able to strategize and provide specific approach in different scenarios (that would be again MD files) – you will enjoy the end result.
Learn what LLMs are, LEARN how they operate (and yeah, learn the Maths behind it – it is not easy but it ain’t Einstein’s field equations either), learn about their fundamental limitations. Then you won’t be repeating shit you don’t understand simply because another dumb ass said it (which is bad enough – spreading the opinion of retards does make one a retard squared).