Vibe Coding Notes

I’ve been playing with “vibe coding” lately. You know the vibe: you describe what you want, the model writes the code, and you pretend you are a calm, visionary product person.

It’s honestly impressive.

LLMs already changed how engineers work. Not just faster typing. It’s more like… the boundary of what one person can do got wider.

But when I tried using it for my research projects, I realized something important. It is much healthier to treat the AI as a junior dev. Ask it to write a specific, small function. Do not ask it to refactor a huge chunk of code. That is unsafe. Also, reviewing a mountain of AI-generated code is a mental burden. The trick is knowing exactly what small tasks to delegate to maximize your own efficiency.

Some people say starting a project from scratch is the best way to “vibe code.” So I tried it on a few small, independent projects with clear requirements, mostly using Cursor.

The Rule: I tried to write as little code as possible. I only described requirements or bugs. I let the LLM build and debug freely.

Here’s what I felt.

Projects Tried#

The Resume Refactor#

Repo: yuan2-resume.

This started because I wanted to organize my past work and plan my future. Then I thought: fine, let’s update the resume. I found a template I liked: yuan-resume. It looked great. But in real use, I hit alignment issues. Some macros also felt too specific. Not very reusable.

That sounded like a perfect vibe-coding task: clear goal, real constraints, lots of boring details.

My original dream prompt was basically:

Based on this template: reduce magic-number spacing hacks, use modern LaTeX packages, and expose clean, general interfaces.

The AI only partially worked. As the “vibe coding” went on, the code started to bloat. It went against my goal of simplicity. Also, asking AI to handle “abstraction” is tough. It rarely matches what you have in mind.

I actually finished the core code by chatting with Gemini and GPT. I defined the interfaces myself. I described the behavior precisely. Once I gave specific constraints, the AI filled in the details nicely.

The AI for CryoEM Project Page#

Site: cryofm.

This was for a recent research project. And web dev is not my comfort zone. Social media says AI is great at building websites. Annoying statement. Also… kind of true.

I chose the Astro framework and a portfolio template ricoui-portfolio. I opened Cursor and transformed into an annoying Product Manager. The AI became my infinitely patient programmer.

For many edits—text, layout, theme colors, Cursor could understand the repo, locate the right files, and make changes. That saved a lot of time.

But I also hit “not-AGI-yet” moments.

One example: I asked it to format a pseudo-code block in the blog. The line numbers and code indentation got messed up. I tried to explain the desired alignment again and again. No luck. I even used a screenshot. Still no luck.

Then I switched approach and told it:

Do this as a two-column table. First column is line numbers. Second column is code.

Boom. Solved.

Rebuilding this Personal Site#

This was inspired by an article praising Claude’s frontend skills¹. I decided to rewrite my personal site using the Pure theme in Astro.

I used GPT-5.2 Codex (via Cursor) to help with the design. It was surprisingly good. It broke away from those boring “safe design choices” and gave me the pretty homepage design.

But it still had moments where it clearly wasn’t AGI. For example, I wanted my publication cards to match the background style of my blog cards (transparent, gradient). I asked for a 1:1 replica. Multiple times. Still not quite right.

In the end, I asked it to point me to where each effect was implemented. Then I manually copy-pasted and aligned it myself.

Lessons Learned#

LLMs have definitely boosted my solo productivity. But to use them well, I learned a few things:

1. Be Precise: You must describe your requirements clearly.
2. You Are Still the Brain: It isn’t true intelligence yet. Human problem-solving and technical architecture are still efficient.

And yes, I’m still amazed by how fast this technology evolves. We’re creating more and more content to feed the next model. Meanwhile AIGC is already messing up parts of the internet. What’s the lasting human value in that world?

Honestly… I’m not sure yet.