AI coding tools often default to generating UIs using React, especially with Server-Side Rendering (SSR), even for simple projects. This makes debugging difficult for casual users. Requesting an option or preference for generating UIs with vanilla JavaScript, CSS, and HTML for simpler use cases.
"Don't write any code yet" is a surprisingly important instruction when working with AI coding tools. Not all tools respond to it the same way. Whenever I begin writing code with an AI tool, whether starting a new project or iterating on something existing, I usually want to first discuss the requirements, architecture, and overall implementation strategy. For example, is this a quick throwaway project or the start of something bigger? In my experience with several AI-assisted coding tools, many are far too eager to start generating code prematurely. ⭐️ Best: ChatGPT. I've found ChatGPT to be the most nimble in going between talking about code and writing code. We can start by discussing a project at length, then iterate on the codebase, and then return to discussing a new feature or a refactor. It's also excellent for hobbyists or semi-pro technologists. Its data analytics features are especially user-friendly. And if you're working on a simple web app, it will gladly keep everything in a single file with vanilla JavaScript and a simple structure that can be copied into JSFiddle, uploaded to a personal web server, or hosted in whatever simple way you prefer. ↗️ Improving: GitHub Copilot, Claude Code. GitHub Copilot was the original "eager coder", spitting out code at the earliest possible moment. As it has matured, and with the release of Agent Mode, it’s now entirely possible to discuss requirements and implementation details at length before it starts churning out code. Claude Code also excels at this, arguably even more so. However, these are professional-grade tools. At a minimum, they require a working knowledge of how to effectively use an IDE. That’s a significant barrier for many and a non-starter when teaching folks with no intention of becoming software engineers (or kids who are just beginning to learn computational literacy). ❌ Needs Work: Replit, Claude Artifacts. Unfortunately, the two most seemingly user-friendly tools are also the least effective at holding back from writing code right away. In my experience, once Claude Artifacts senses that you want to create software, it will begin generating code immediately and won't stop tinkering with it at every opportunity, no matter what you tell it. Replit, too, is currently designed to build early and often. The gracious interpretation of this is that these tools want to quickly show you working prototypes that will help you iterate. Maybe this will be true in due time. But today, these tools frequently paint themselves into a corner: premature assumptions about requirements, poor technical decisions (sometimes using the proverbial cannon to kill a fly), and bug-ridden code that dead-ends the project. Too often, the only way out is a new conversation to try again. There are many other tools out there I haven't tried, like Cursor and Bolt. What's been your experience? #AICoding #PromptEngineering #SoftwareDevelopment #VibeCoding