How I use AI to code

AI coding agents are the hype of the moment - and how could they not be? They promise to be smart, to build fast, to iterate even faster and to deliver working and elegant solutions to all your problems.

Vibe-coding has entered all our lives, in a way or the other, and everyone talks about it as the current big thing in tech and the ultimate bridge between technical and non-technical people when it comes to bringing ideas to life.

In this context, I asked myself: why don’t I write about my general coding flow with AI and the tools that I’ve been using?

As a disclaimer: I am not an AI power user, I don’t have highly optimized coding workflows or dozens of sub-agents burning hundreds of thousands of tokens every day for me. I’m just a girl who happens to have found a way to integrate AI in her flow, and wants to share advice coming from her practical experience: with this being said, let’s dive in!

My Dev Flow with AI

Let’s start from the beginning, from the idea: when I have an idea for something, or when I have to write some code that I do not immediately know how to implement, I first of all stop and think. I think about what I am trying to achieve and what I need to achieve it, and then I start my research.

Generally I search the web, skim through GitHub repos from other people to see their implementation or to find useful resources for my project, and finally, if I can’t find anything useful on that side, I ask AI to give me hints on where to look, what to use or how to design the project.

Once I’m out of this phase, the real building starts: building for me it’s first of all sketching ideas, designing the project on a high level, creating the backbone of it.

I generally start from the backend and then move onto the frontend: this is simply because, being a backend-first engineer, it is easier for me to design the frontend around the needs of the backend.

Once the first sketch of the project is implemented, I test it and see if I like the results.

This test phase is generally the one where issues and ideas on how to improve the project emerge - and generally these things require more complex implementations.

Whenever there is a more complex implementation, I first of all think about it and spend some time designing it in my head, then I take a stab at it. It is at this point that generally AI comes in: once I am done with my implementation, I ask AI to evaluate it and tell me whether or not it works, if there is something I can do better and if there are any errors. I don’t have any magical prompt - I just copy-paste the code, give a bit of context on what I am trying to achieve and ask some plain questions like: “Is this implementation correct?”.

The most important thing for me, when using AI, is that I do not offload the ideation/design phase to AI: I have a clear idea of what I am building, what it should look like and what outputs are supposed to come out.

Once I integrate AI’s code into my project, I test it: if there is an obvious flaw I can fix, I do it, otherwise I get back to the AI with the bug (generally an error traceback and/or a screenshot of my app) and iterate on the solution.

This is pretty much it: once I’m satisfied with the code, I push it to GitHub and, occasionally, deploy it to a live demo/prod application.

I mentioned “AI” several times and I kept it generic on purpose, because now I want to give you a quick overview of the AI tools I use to assist myself with coding.

My Tools

Claude

I use Claude (the LLM app, not Claude Code) daily: Claude is my go-to for any programming question, but I feel like it is especially valuable for languages like TypeScript and Go, and for complex python questions.

I generally give Claude my code and ask for clarification, error explanation, improvements or design tips (especially for frontend design): I generally iterate two or three times until I get to the code I want, unless it is a simple task (like centering a div) - in that case one-shot generation works more than fine.

GitHub Copilot

I use GitHub Copilot inside of VS Code, mostly to refactor my code when the refactoring is just mechanical (like changing the arguments of a function in all the occurrences in my script) or when I need docstrings/annotations for my functions. I tend to use it mostly with Python, but sometimes I also use it for TypeScript.

My go-to model for GitHub Copilot is GPT-4.1 (I’m a fan girl, I can’t do anything about it🤷‍♀️)

v0

v0 is a blessing for my frontend and full stack development.

I can easily say it feels like the most powerful assistant I’ve ever worked with in terms of designing, debugging or correcting frontend apps: even better than Lovable (although their scopes are slightly different).

I use v0 especially for UI/UX design questions: I sketch the frontend with all the components I need to fit the backend, then I go to v0 and ask for help on styling. Also, it is an invaluable resource for developing with NextJS in a React environment using shadcn/ui: every time I have a doubt related to a component or how it works, v0 generally knows the answer.

Gemini and ChatGPT

My use of Gemini and ChatGPT is near-zero for coding-related task, but still I get to use them sometimes, especially as a replacement of Claude when I run out of credits.

ChatGPT, in my experience, offers convoluted solutions and extremely verbose explanations that oftentimes I lack the time or the will to read, which in turns leads to an impoverished DevX because most of the times ChatGPT’s solutions can work, you just have to carefully read the instructions. Still, I mostly use it for simple-to-mid-level problems, and especially for python.

Also Gemini has the flaw that it produces very long instructions along with the code that should help you: I have to say that a lot of times Gemini fails to understand my problem, or to solve it, but it is good for ideation or for sketching things from scratch, and I used to ask for its help a lot some time ago to get started with full-stack applications. Today I mainly use it as a replacement for Claude on TypeScript/Go questions.

These is basically how I use AI for coding: I know many things are easy to disagree with, and if you have other opinions please feel free to share them with me on one of my social pages!