Back in May, I shared how I built my dream GTD app in just a couple of hours using Replit’s AI agent. This got quite a few in my network, excited. “Tell me more about this app building thing Paul”, “Can I use it to rebuild my …..” People are generally excited about the possibilities of AI-powered development. I promise to do a separate blog specifically answering those questions.
In the meantime, three months after building my GTD app, (and I’ve been away in Europe on vacation for a significant chunk of that time) I want to share what I’ve learned from actually living with an AI-built app and my ongoing experience with Replit.
The Reality Check: Does It Actually Work?
The short answer is yes. My custom GTD app has become an integral part of my daily workflow, and I’ve become somewhat addicted to iterating on it and improving it. Unlike the countless off-the-shelf solutions I’ve abandoned over the years, this one is sticking because it does exactly what I need and when it doesn’t, I fix it.
The true differentiator of my GTD app’s is how it surfaces the next action from each of my projects. This is tailored to the way that I work. I open it to see what matters most across my various projects, without getting lost in endless task lists. What is the next action across each my projects, get that done and keep the momentum moving forward. Is my app perfect, No, but most importantly its simple, it’s mine, and it works.
The story I want to share with you in this post is about my continued understanding of using Replit’s tool to build and iterate on my app. This has been as enlightening about the current state of AI development agents as it has about building productivity apps.
Two AI Developers Walk Into a Codebase…
One of the most interesting discoveries has been working with Replit’s dual AI approach. Replit has Agent which is designed to create full-stack apps from scratch, handle complex features and integrations, and manage environment setup and Assistant (which came first) focused on quick edits, chat-based assistance, and smaller modifications to existing apps.
I’ve set my Assistant to use Claude (you can choose between Claude 4.0 Sonnet or GPT-4o) while the Agent based on my experience is using a different model, I am unsure if this is Replit’s own proprietary model, or a wrapper on top of something like GPT. In any case, it’s like having two expert developers at your disposal, each with their own personality and strengths.
The Agent is fantastic for architectural decisions and complex feature implementation—it thinks big picture and can scaffold entire functionalities quickly. Claude, through the Assistant, excels at refinement, debugging, and those careful, methodical improvements that polish an app. With that said, I’ve used both in various ways, there is quite bit of overlap. Sometimes one works better for a particular task, sometimes the other. Having both available feels like the beginning of something much larger in terms of how we’ll collaborate with AI in development.
Some Things Never Change: Niggly CSS
Here’s what surprised me a bit, 20 years since I last coded professionally, I’m finding some of the trickiest issues are still formatting and CSS-related problems. Even our AI overlords struggle with this.
There’s something comforting about watching an AI agent—capable of reasoning through complex algorithms and architectural patterns—get completely stumped by the eternal frustrations of web development. CSS remains the great equaliser!
When AI Gets Stuck: Loops, Baselines, and Human Intervention
Perhaps the most fascinating aspect of working with Replit’s agents has been observing how they can get caught in loops—and how much this resembles the experience of being a human developer. Sometimes an agent will chase its tail trying to fix an issue, implementing and re-implementing the same approaches with slight variations, much like I’ve done countless times in my career.
The solution is surprisingly familiar: recognise you’re stuck, restore to a previously stable baseline, and try a completely different approach. What’s interesting is how easy this is with AI agents—you can literally roll back and restart the conversation from a different angle.
This has me wondering whether this kind of visibility into “stuck states” and easy recovery mechanisms might be valuable for traditional development teams. Maybe AI could help facilitate this kind of meta-awareness: “Hey, your team has been working on this issue for three hours and trying variations of the same approach. Here are three completely different ways to think about this problem.”
The Unexpected Evolution
What I didn’t anticipate was how naturally my app would continue evolving. Major feature requests that would traditionally require planning meetings, sprint allocation, and weeks of development became afternoon conversations. This isn’t just about speed—it’s about removing the psychological friction that prevents us from improving our tools.
One aspect I hadn’t fully appreciated in my original post was how Replit’s platform approach has fundamentally changed when and where I can develop. I now have their desktop IDE for serious development sessions and their mobile interface for quick tweaks and iterations. This means I can literally continue building wherever I am.
Stuck in a waiting room? I can refine my app’s user interface. Commuting on the train? Perfect time to ask the AI agent to implement that feature I thought of during my morning coffee. Traveling? No problem—everything lives in the cloud, accessible from any device.
This isn’t just convenient; it’s transformative. The traditional model of development required being at your development machine, with your specific setup, tools, and environment. Now, the barrier between having an idea and implementing it is just having an internet connection. I’ve made meaningful improvements to my GTD app from coffee shops, airport lounges, and even from my phone while walking the dog.
The psychological impact of this accessibility can’t be overstated. When you can act on inspiration immediately, you actually do. Ideas don’t get lost in the gap between conception and the next time you sit down at your development machine.
The Professional Reality: Where Replit Hits Its Limits
While my personal GTD app thrives in Replit’s environment, I’ve started pushing against some boundaries that reveal where these platforms currently sit in the development ecosystem. Setting up proper Dev/Staging/Production environments isn’t straightforward out of the box, which I believe limits the use cases where Replit can be a complete solution.
This isn’t a criticism—it’s more about understanding where we are in the evolution of AI-powered development platforms. They’re phenomenal for rapid prototyping, personal projects, and proof-of-concepts. But the bridge to professional software development workflows still has some gaps.
I suspect this is where we’ll see significant innovation in the coming year. The ability to rapidly prototype with AI is clearly here; now we need the infrastructure to seamlessly transition those prototypes into production-ready applications with proper deployment pipelines, environment management, and scaling capabilities.
What This Means for the Future
My experience points to several trends that I think will define the next phase of software development:
The Rise of Personal Software
We’re moving toward a world where having custom software for your specific needs no longer is a luxury reserved for developers, Replit and tools like it are federating this capability to all of us. Why struggle with software that almost fits when you can build software that fits exactly?
Multi-AI Development Teams
The future might not be about finding the “best” AI for development, but about orchestrating teams of specialised AI agents. Different models for different strengths, working together on the same codebase. It is almost like development pairing, but Replit hasn’t quite got the AIs communicating with each other yet. You can interface with each of them, and they can both work on the same codebase, but they can’t talk to each other or interrogate your interactions with the other AI. I wonder why that is, and what the limitation is preventing it.
The New Role of Human Developers
Rather than being replaced, human developers might become conductors—directing AI agents, recognising when they’re stuck, and providing the strategic thinking that bridges between rapid prototyping and production-ready software.
Infrastructure as the New Frontier
The bottleneck is shifting from “can we build it?” to “can we deploy, scale, and maintain it?” The next breakthrough might not be better AI agents, but better AI-native infrastructure.
What’s Next
My GTD app continues to serve me well and evolve with my needs. But more importantly, this experience has opened my eyes to what’s possible when the friction of software creation disappears. I’m now thinking about several other small tools, each perfectly tailored to specific needs, each built in hours, days or weeks, rather than months and years.
I’m also planning to explore Replit’s environment limitations further and document what it takes to bridge from rapid AI-assisted development to production deployment. There’s a fascinating problem space emerging around AI-native environment workflows that deserves deeper exploration.
The Real Lesson
The real lesson from my continued GTD app journey isn’t about productivity or even AI—it’s about agency. For the first time in a long time I have true agency over my digital tools. But more than that, I’m witnessing the birth of a new kind of human-AI collaboration in software development.
The age of bespoke software for everyone is beginning. And honestly? Even the CSS headaches are worth it.
My GTD app is still running at https://realvelona-gtd.replit.app/ if you want to try it out. As always, it’s a prototype—use at your own risk, but feel free to be inspired by what’s possible.
Leave a Reply