From Mobile‑First to AI‑First: Why I’m Re‑Tooling My Entire Dev Stack

A feature that previously took a few hours was delivered in minutes after I asked my AI code assistant for help. Could I have done it myself? Yes, but what’s the point if AI helps me do it faster, without unnecessary boilerplate?

That’s when I knew I was no longer a "mobile developer" but an AI-first developer.

I started using AI in 2024. Initially, hallucinations were common, mostly due to mismatches of old library versions, nonexistent functions, or broken code. But things have dramatically improved since then.

Why the Ground Just Shifted Under Us

Since early 2024, we’ve seen massive improvements. We now have AI-powered IDEs, capable of modifying code automatically without tedious copy-and-paste from tools like ChatGPT. The most transformative shifts have been:

1. Web Search: AI’s ability to fetch up-to-date information on libraries, SDKs, and read documentation seamlessly.
2. Larger Context Windows: Allowing us to include our entire codebase, README files, examples, and system prompts within a single AI context.
3. Improved Models: Especially notable are Claude Sonnet 3.7 and the latest GPT models (4.1, o3), which offer significantly fewer hallucinations and higher accuracy.
4. MCP Tools: AI can now directly read and write files, execute shell commands, interact with databases, and understand database schemas automatically.

Recent benchmarks clearly indicate the advancements: Claude Sonnet 3.7 ¹ or GPT-4.1 ² are killing it in coding benchmarks.

What Makes a Codebase AI‑Friendly

There has been a significant shift in what AI can and can’t effectively handle. Using languages, frameworks, and build tools that expose rich AST (Abstract Syntax Tree) and LSP (Language Server Protocol) metadata is essential. Languages like TypeScript paired with frameworks like React and Tailwind provide the ideal environment for AI-driven development, allowing AI to effortlessly generate robust and maintainable code.

Attempting to use AI for niche or highly specialized technologies that aren’t widely documented (trained on) or commonly used often yields poor results. AI thrives in widely-adopted ecosystems supported by extensive open-source contributions and online documentation.

Thus, when beginning a new project with AI assistance, it’s crucial to choose a technology stack that plays to AI’s strengths:

If your stack looks like the left column, AI assistance will stumble.

My New Workflow (Steal It!)

Choose an AI-friendly stack and start with simple, direct prompts. There’s no need for elaborate politeness or overly detailed introductions. Just clearly state the goal and required tech stack in concise bullet points.

Modern AI models are already well-trained to handle initial project setups, including configuration choices, dependency installation, and boilerplate code generation.

For instance, with a clear prompt like this:

Goal: blog website
- Stack: TypeScript, React, Next.js, Tailwind
- Package manager: Bun.js
- Additional tools: shadcn-ui, react-hook-form
- Pages required: /about, /posts, /contact

the AI will handle the majority of the initial setup, leaving you to focus on adding specific features, functionality, and necessary refactoring.

The Road Ahead — Code Less, Build More

2025 has confirmed what many of us felt in our gut last year:

AI isn’t an optional accelerator any more, it’s the default gearbox of modern software engineering.

References