For months, I've been coding by hand the old-fashioned way

I've gotten used to producing results quickly with coding agents, but when it's unclear what I actually want, the agent ends up making assumptions on my behalf, which reduces how much I learn and dulls my sense for code — that's the problem awareness this starts from.

So for the past few months, I've deliberately been writing code by hand, spending time re-internalizing the codebase physically and rebuilding my problem-solving instincts. The goal isn't simply to recover productivity, but to rebuild the foundation needed to use agents better.

At the center of this process is Recurse Center. There, I relearned LLMs by implementing a Transformer from scratch, and completed the Stanford CS336 assignments without LLM help. I built a tokenizer by hand, constructed a GPT-2 style architecture in PyTorch, and ran ablation and hyperparameter experiments based on Tiny Stories and OpenWebText.

After that, the plan is to move on to tasks like GPU profiling, implementing FlashAttention2 in Triton, and model post-training. At the same time, to get better at Python, I've been writing a lot of small agents and neural networks by hand myself, and I say the biggest help came from pair programming with people who've been using Python for over 10 years.

At RC, beyond Python, I also experienced Unix, Vim, Clojure, mob programming, and CTF Fridays. I especially emphasize that working with limited tools instilled the habit of thinking before typing code and building the program's structure in my head first.

I also share the observation that the better a programmer is, the higher their AI usage skill tends to be. Only with deep understanding can you leverage agents and LLMs as a bigger lever — leading to the conclusion that, in the long run, hand-coding practice actually boosts your ability to use agents.

Even if I don't finish all my goals in the remaining time, I emphasize that the value of RC lies not in clearing a checklist but in the act of spending time coding itself. In the end, this is a firsthand account of why, even in the AI era, it still matters to move your own hands and relearn concepts, tools, and layers of abstraction.