The Tools

I What Is Vibe Coding?

In February 2025, Andrej Karpathy — one of the founding members of OpenAI and former head of AI at Tesla — posted a tweet that gave this entire movement its name:

There's a new kind of coding I call "vibe coding," where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. [...] I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works.

— Andrej Karpathy, Feb 2025

The tweet went viral because it named something people were already doing. Engineers, designers, and — increasingly — product managers were discovering that AI tools had crossed a threshold: you could describe what you wanted in plain English and get working code back. Not perfect code. Not production-ready code. But workingcode, in minutes instead of weeks. By late 2025, Karpathy himself reportedly hadn't typed a line of code in months — the person who coined the term was living it full-time.

This series takes a clear stance: use AI to build, but stay in control. The "fully vibing" approach works for throwaway experiments. For anything with users, data, or consequences, you want the directed collaboration end of the spectrum. The goal isn't to eliminate thinking — it's to eliminate the gap between thinking and building.

II Why This Changes How You Build

You've scoped the feature. You've debated edge cases. You've written it all down. And then you wait — for the build to come back, for the review cycle, for the version that doesn't quite feel right in practice. Days. Sometimes weeks. By the time you see a working version, the market has moved, requirements have shifted, and the thing you specified doesn't match what you actually needed.

Vibe coding collapses that feedback loop. Not for everything — you're not going to build a distributed payment system by talking to an AI for an afternoon. But for a surprising number of workflows, the gap between "idea" and "interactive prototype" can shrink from weeks to hours.

Working prototypes beat specifications. A clickable prototype communicates intent better than any document. When you can put a working thing in front of someone — something they can click, break, and react to — the conversation changes. You stop debating abstractions and start refining reality.

You carry context that's hard to transfer.You've talked to users. You've read the support tickets. You know whythe feature matters in a way that's hard to fully capture in writing. When you're the one prompting the AI, that context flows directly into the code without lossy translation.

AI-native teams are already forming.Companies are discovering that the most effective AI products come from teams where everyone can prototype with AI tools. This isn't about replacing specialists. It's about arriving at the decision point with a working demo instead of a slide deck. Andrew Chen predicts that coding will sit alongside email, slides, and word processing in every white-collar job description within 18 months — not as a specialization, but as a baseline expectation.

III Replit

Replit is the tool that makes vibe coding feel the most like magic. You open a browser, describe what you want, and watch it build — the code writing itself in one panel while a live preview updates in another. No terminal. No local setup. No "works on my machine" problems. Everything runs in the cloud, and you can deploy to a public URL with one click.

Replit's AI agent is genuinely good at scaffolding full-stack applications. Tell it "build a task management app with user authentication and a Postgres database" and you'll get something functional in minutes. The agent understands frameworks, can install packages, debug errors, and even fix its own mistakes when the preview breaks.

Where Replit shines:

• Zero setup. Browser-based. Works on any machine. No installs, no configuration, no environment debugging.
• Visual feedback loop. You see changes in real-time as the AI writes code. The preview panel updates live, so you know immediately if something looks wrong.
• One-click deploy. Your project goes from code to public URL instantly. No CI/CD pipeline to configure, no hosting to manage.
• Collaborative. Share a link and someone else can see your code, your preview, and your deployment. Great for demos and stakeholder reviews.

Where it falls short:

• Less control over architecture. The AI makes structural decisions for you. When those decisions are wrong, unwinding them is harder than it would be in a tool that gives you more control.
• Complex backends get messy.Simple CRUD works great. Multi-service architectures, complex business logic, or anything requiring specific infrastructure patterns pushes against Replit's strengths.
• Vendor lock-in.Your code runs in Replit's environment. Migrating to a different hosting provider later means extracting and adapting everything.

IV Claude Code

Claude Code works differently. Instead of a browser IDE, it lives in your terminal. You open it in a directory — any directory, on your machine, with your files — and start a conversation. It reads your existing code, understands the project structure, and makes changes in place. There's no separate environment, no cloud workspace. It works on your stuff.

This matters more than it sounds. When an AI tool has full context of your actual codebase — your folder structure, your naming conventions, your existing patterns — the code it writes fits in. It doesn't scaffold a new project from scratch; it extends the one you already have. For anyone building on top of an existing codebase, this is the difference between useful output and output you have to rewrite. Boris Cherny, the head of Claude Code at Anthropic, frames it bluntly: coding as a bottleneck is effectively solved — the real constraint has moved upstream to design, architecture, and intent.

Where Claude Code shines:

• Deep context. It reads your entire project. When it generates code, it matches your existing patterns, imports the right modules, and follows conventions it observed.
• Works on real projects. Not just new scaffolds. You can point it at a mature codebase and ask it to add a feature, fix a bug, or refactor a module.
• Git integration. It works with your version control. You can review diffs, commit selectively, and maintain a clean history — the same workflow engineers use.
• Iterative refinement.The conversation persists. You can say "that's close but move the sidebar to the left" and it remembers everything from the previous turns.

Where it falls short:

• Requires terminal comfort.If you've never opened a terminal, there's a learning curve. It's not steep, but it's there.
• No built-in deploy. You write the code locally. Getting it live means pushing to a repo and setting up hosting separately — Vercel, Netlify, GitHub Pages, whatever your stack uses.
• No visual preview. You see code changes, not a live preview. You need to run the project yourself to see the results.

The honest assessment: Claude Code has a higher floor (you need some technical setup) but a higher ceiling (you can build more complex, production-grade things). Replit gets you to "wow" faster. Claude Code gets you to "shipped" more reliably.

V The Landscape

Replit and Claude Code aren't the only options. The AI-assisted building space is moving fast, with new tools launching monthly. Here's how the current leaders compare across the dimensions that matter most.

A few patterns emerge from this comparison. If your primary output is a visual interface — a landing page, a dashboard, a design prototype — Lovable or Replit will get you there fastest. If you're working on something with real backend complexity or need to integrate with existing code, Claude Code or Cursor are the better bet. And if you're comfortable in VS Code already, Cursor gives you AI assistance without changing your workflow.

The right answer isn't "one tool." Most practitioners end up using two or three depending on the task. Replit for fast demos, Claude Code for serious building, Lovable when design matters most. Think of it as a toolkit, not a religion.

VI What's Next

You've got the map of the tooling landscape. You know what each tool does well and where it breaks down. But having the right tool is only half the story — the other half is knowing how to use it.

In Part 2, we'll cover the craft of vibe coding: how to write prompts that actually produce what you want, how to iterate without losing context, and the workflow patterns that separate a frustrating afternoon from a productive one. The tools are the instrument. The craft is learning to play.