Vibe Coding, Demystified

{% card %} **TL;DR -**Vibe coding is the practice of building real software by talking to an AI, in plain English, instead of writing every line by hand. The whole stack a beginner needs fits into four roles: a conversational AI tool for the rough draft, an AI-powered editor for the real work, a version control service to save progress, and a deployment service to put the project online. The human still picks what to build, judges whether the result is any good, and decides when it ships. The on-ramp is shorter than it has ever been.{% endcard %}

Vibe coding has become the shorthand for a way of building software that did not really exist three years ago. A person opens a conversation with an AI tool, describes what they want, gets working code back, and ends the session with something live on the internet. No bootcamp. No three-hour Homebrew install. No semicolon hunt at 2am.

This piece is for the software creator: the person who wants to build something useful, not someone trying to become a senior engineer at a megacorp. The advice differs depending on which path a learner is on, and the previous piece in this series sketched both. What follows is the practical shape of the software creator path, the tools that show up in it, and the moments where the human in the loop still really matters.

What an Actual Session Looks Like

The best way to understand vibe coding is to walk through a real one. Let’s examine a site revamp: An experienced developer helped their partner, a cinematographer, replace his Squarespace portfolio with something built from scratch. He had no developer background.

The flow looked something like this:

She pointed him to Google AI Studio and told him to prompt whatever he wanted. He typed something close to “build me a portfolio website for a cinematographer.” A first mockup came back. He kept iterating in conversation until the layout felt right.
Then he downloaded the code. He opened it in an AI-powered editor, which lets a user see the actual project files alongside a chat window. Then an AI assistant inside the editor handled the next round of work: pulling his real bio and project list out of the old Squarespace, making styling changes, and answering his questions along the way.
When he was ready to make it live on the internet, he connected the project to GitHub for version control, then used a hosting service to make it live. He already owned the domain. He asked the assistant which fields to update at his domain registrar, and it walked him through them.

The whole project took him a few weekends of conversation to get to a real website. No tutorial completed start to finish.

When a vibe coding session goes well, it looks just like this. The human is making decisions about taste, structure, and intent. The AI is handling the parts that used to take three hours of YouTube and a sympathetic friend.

The Tool Stack, by Role

The tool landscape can feel overwhelming when seen all at once, partly because the same task can be done in five different products. The simpler way to navigate it is by job. A vibe coding workflow generally pulls one tool from each of these four categories:

Conversational Sketchpad: This is a general AI tool where the first version of an idea gets generated. Google AI Studio, Claude, and ChatGPT all work for this. Some people stay here for the whole project. Others use it just for the rough draft.
AI-powered Editor: This is where the code gets edited with help from an assistant that can see all the files at once. Cursor and Google’s Antigravity are two examples. VS Code with the Copilot extension is another widely used option. These editors are friendlier than dropping straight into a terminal, because the user can see the files and the changes happening in real time rather than typing blind into a black window.
Version Control: We’ve discussed this before. GitHub is the default for most people, and it deserves its own piece in this series. GitLab and Bitbucket exist as alternatives. The function is the same across all three: it tracks every change to the project, so nothing is ever truly lost and rolling back to a working version is always one click away. Think of it as Google Docs history for code.
Deployment Service: This is where the finished thing actually lives on the internet. Cloudflare Pages and Render show up often. Vercel and Netlify are also widely used. Most of these connect to GitHub directly and update the live site automatically whenever the code there changes.

A workflow can use one tool per row, four total. That is really all the infrastructure that your first project needs.

A Note: The differences between the leaders in each tool category are small enough that a beginner will not feel them. Familiarity, free tiers, and whatever a friend already uses are reasonable tiebreakers. The right tool is mostly the one that actually gets used.

Where the Human Still Matters

Vibe coding sometimes gets caricatured as the user typing “build it” and the AI building it. The reality is more collaborative, and the moments that require human judgment tend to cluster in three places.

To begin with, the AI cannot read your mind, so the quality of the first prompt shapes everything that follows. “Build me a portfolio site” is a workable starting point. “Build me a portfolio site for a cinematographer with a dark theme, full-bleed video on the landing page, a project grid that filters by year, and a contact form that emails me directly” gives the AI much more to work with and produces a much more useful first draft. Specificity is the entire skill.

Next is at the breakage point. AI assistants are very good at writing code that works on the first try most of the time. They are less reliable at catching when something subtly does not work. Examples include: A login that lets anyone in; a database connection visible to the public; a form that quietly fails on phone keyboards. These failures look like working code…until they don’t. Asking the AI to audit its own work, after the fact, with prompts like “what would a senior developer change about this?” or “walk me through the security of this database connection,” catches a good portion of these issues before they hit anything real.

Finally, AI can only offer you options – it can’t make decisions for you. What gets built next? What stays in scope? What is good enough to ship? You have to decide what matters.

A Real First Step

Here’s your next move: Open a free account at any of the conversational AI tools mentioned. Ask it to build a simple version of something small and personal: a one-page site, a photo renamer, a tracker for whatever the reader is currently tracking on paper. Talk to it the same way one would talk to a patient friend. Ask follow-up questions when the response does not make sense. Save the project to GitHub when it works. Push it to the internet when it feels real.

The shape of software development has changed. The on-ramp is dramatically shorter, and the number of people who can credibly call themselves builders is about to grow by an order of magnitude. For someone just starting out, the question worth dwelling on now is what to build. The other questions sort themselves out from there.

MLH hackathons are a good place to bring that first project into a room of people doing the same thing. DEV is a good place to write about what got built once it ships. The build happens privately. The learning happens in public.

Frequently Asked Questions

{% details Is vibe coding actually coding? %} Depends on the definition. The person doing it ends the day with working software that runs on real infrastructure, which is the part most people care about. The activity looks more like directing than typing. Many working developers now describe their own day-to-day this way.{% enddetails %}

{% details Do I need to know any code before I start? %} No. The cinematographer in the example above did not. What helps is being specific about what is wanted and patient about asking follow-up questions when the AI does something unexpected. Knowing a little code makes it easier to audit the output later, which matters more for anything handling real user data.{% enddetails %}

{% details Is the code any good? %} It is usually functional. It is sometimes great. It is occasionally subtly wrong in ways that matter. Code that handles passwords, payments, or personal information deserves a second look from someone who knows what to look for. Code that renames files in a folder probably does not.{% enddetails %}

{% details What happens when I get stuck? %} Most beginners stall in one of two places: a feature that will not work no matter how it gets reprompted, or a deployment step that throws an error message in unfamiliar language. The fix for the first is usually to restart that part of the conversation with a more specific prompt. The fix for the second is to paste the exact error message back into the AI and ask what it means. Search engines, the DEV community, and an MLH event are also genuinely useful at this point.{% enddetails %}

{% details Is this how real developers work now? %} Many of them, yes, at least for parts of the job. Senior engineers tend to use AI as a collaborator for unfamiliar territory and a multiplier for familiar territory. The skill that distinguishes them from beginners is knowing when to trust the output and when to dig deeper. That skill is learnable, and it develops through practice.{% enddetails %}

{% details How much does vibe coding cost? %} Most of the conversational AI tools have generous free tiers that are enough for a first project. GitHub is free for public repositories and most personal use. Cloudflare Pages, Render, Vercel, and Netlify all have free tiers that cover small sites. A custom domain costs roughly ten to fifteen dollars per year. A learner can ship a real first project for the price of a domain name and nothing else.{% enddetails %}

{% details Should I still learn to code the traditional way? %} That depends on the goal. Someone aiming for a software engineering career will benefit from foundations: a real language, the basics of how computers actually run code, some experience reading code written by other people. Someone who just wants to build a thing that solves a problem in their own life can probably skip that, at least at first. The previous piece in this series breaks out both paths in more detail.{% enddetails %}