// LATEST

Stop Treating AI as a Wish-Granting Machine: On How to Start a Real Vibe Project #1

We've now entered the great VibeCoing era: you can build a project with a single sentence—just keep making wishes and throwing requirements at AI, and the project gets done.

But do you really know how to maintain a VibeCoding project like this long-term?

Here are a few tips I've summarized.

1. Break Down the Project — Version Control Thinking

Don't expect AI to produce a finished product in one go, and don't give AI too many requirements at once; instead, move forward step by step.

Instead, you need to break down the project boundaries.

I highly recommend systematically learning Git's version control mindset. There are mainly two points.

  1. branch: Split requirements or goals into branches, with each branch focused on doing one thing. For example, adding a settings screen to the software. This kind of boundary isn't meant to constrain AI—it's meant to constrain yourself. Through this constraint, you make yourself focus on only one feature each time you code, instead of making wishes all over the place—when Vibe makes you feel omnipotent, what matters isn't what features to add, but what should be done and what shouldn't.
  2. commit: Break each round of work into finer pieces, split into individual commits. On one hand, this prevents AI from messing around and breaking the project; on the other hand, it gives you control over the whole project and lets you know what AI did.

You can do these two things yourself, or let AI do them autonomously—we'll cover that later.

2. /plan

A very important step in moving from wish-based programming to actually building a project.

When you're working on a new branch, don't just blurt out one sentence. That kind of wish-style gacha programming is very uncontrollable, and AI always churns out crap.

Instead, you should write out a plan: what you're doing this time, what the boundaries are, what tech to use, etc...

But writing these plans is really a hassle. So why not try /plan?

In plan mode, discuss with AI and eventually arrive at a suitable plan.

3. Don't Use a Single Conversation

Many people keep talking in one conversation forever. This is actually a bad habit. A single long conversation has the following problems:

  1. Large context, leading to high Token overhead
  2. Context pollution, which leads to scattered attention
  3. No multi-agent parallel development

So, start a new conversation for every new branch. Also, to prevent the problem of automatic compression halfway through a task, it's recommended to manually compress after an appropriate commit. At around 50%+, you can start considering manual compression at a suitable stage. Of course, the best case is still to finish the entire branch within one context—this is also a reference for how big to make a branch.

One problem with starting a new conversation is that every time, AI has to first read the entire project and understand it. Obviously, this isn't ideal. I'll explain how to solve this in a follow-up post.

No comments yet

Share your thoughts and keep the conversation clear and kind.