My experience using AI as a cheapskate
Yes, you heard it right. I don’t pour my whole wallet into LLM providers. I try to use the free daily tokens from Claude out there. When ChatGPT hits a limit, I jump ship to Gemini. Sometimes I extracted a part of LLM A’s answer and fetched it to LLM B to continue the conversation, and that means accepting the lossy summarization and lacking-the-context responses.
Part of doing so, I admit, is that I’m quite stingy. Growing up Asian makes you perform more with less, and I took it to heart. But there are other reasons.
Not because I don’t trust LLMs’ output, nor do I prefer coding all by typing and pure Google search like in days of yore. I wholeheartedly agree that setting higher output token limit i.e., spending more, will certainly lead to better output. And I have always wanted to use the multi-agent functionality of Claude Code: I heard that we can let it run just for a while and we can have a whole website with backend + frontend up and running, everything with a few good prompts. All in all, I like it a lot, and I want the best out of it even if I have to squeeze in some more cash.
But I have to be honest: Sometimes the AI is just too good, it makes me feel like I’m not learning. The LLMs sometimes use the most cutting-edge features of the language or libraries that abstract all the low-level nitty-gritty stuff, and that makes things way too…convenient. It curbs the curiosity we programmers need to have to understand something. And I sure do not want that. I love deep diving into a details, not inserting another coin into the slot machine, wondering if the output is the solution or more abstraction. Before you chime in to say “Yeah you have to use this magic prompt/tweak this and that/be patient with it/etc.”, believe me, I tried. For me AI is still a hit or miss, with the hit part getting better day by day, but in general we still have a lot of work to be done for it.
And other times it completely derails from what I want. The agents out there are really good, but I have to do a lot of tweaking just to ask a simple question: Changing to the Ask profile, configure the tools so that the LLM does not waste token on searching through the code base but instead just do a simple web search inside the IDE so I don’t have to open a new Google tab (I’m looking at you, Zed), and more typing: “You do not have to write code to examplify the problem, just tell me what is wrong”. The convenience agents bring is magical, but sometimes agents are on the annoying side of things. I also don’t want to switch between this agent to that chat interface back and forth, because that means the context is gone and have to start again from scratch with limited functionalities.
I think all the problems above could be easily resolved if I add a line in the system prompt like “For every line of code you write, add explanation” or “Try to come up with a solution that has does not rely on dependencies and be as minimal as possible”, but the first one might cost a lot more extra tokens and the second on … well … sometimes time is precious and we have to move fast. These excuses are quite weak, I must admit, but they got me dial back my LLM usage a lot of times. Doing so save quite a lot of money for a junior like me too.
Typing out code is educational and above all, fun. Don’t let LLMs take the entirety of that away from you.
