Two projects, two experiences – what happens when AI becomes your coding partner?

Summary

AI coding assistants have become remarkably capable – but how do they hold up in the hands of a real developer tackling real problems? In this article, I share my experiences using two AI tools: Lovable.dev to build a practical web app and Claude.ai to create a small simulation in R. Both projects revealed the promise and limits of current AI systems, and what it feels like to ‘code with a machine.’

The Promise and Reality of AI-Assisted Coding

Over the past two years, AI has crept into every corner of software development. GitHub Copilot completes functions before you finish typing, ChatGPT explains obscure errors, and platforms like Lovable.dev generate entire web apps from a single sentence.

One of my hobbies is beekeeping, and that is where the opportunity to use AI arose. I had two projects where I had limited time and no resources other than myself. Could AI finally automate the tedious parts of development? Or even better, could it code for us? My plan was to:

• Build a website to educate the public about the Asian Hornet and to track where beekeepers have placed traps.
• Create a simulation to illustrate to students how they can use statistics to model situations in bee genetics.

Both projects were modest but real – practical enough to reveal strengths and weaknesses, and distinct enough to show how AI handles both structure and logic. I planned to allow the AI to build the solution with minimal intervention other than prompts, to see if it could code the final solution on its own.

Project 1 – Building a Website with Lovable.dev

The Task

The Asian Hornet has now made its way to Ireland, with nests discovered in Cork and Belfast. While beekeepers understand the risks, this is an issue for everyone. Beekeepers across Ireland organised via WhatsApp and decided we needed a mechanism to track and manage traps placed to monitor the species.

The website’s purpose was straightforward: an educational resource for the public and a system to manage registrations, logins, trap placements, and map displays. I began with a detailed prompt describing what I wanted. Within a couple of minutes, Lovable produced a working React app – pages, navigation, styling – everything. It took several iterations to refine, but the initial output was astonishing.

The Impressive

Lovable handled structure brilliantly. The generated code was modular, readable, and consistent. When I asked for a new piece of functionality, it adapted cleanly and regenerated only what it needed. For basic scaffolding, it was faster than any manual workflow. When I requested integration with a REST API and a MariaDB backend, it complied – though reluctantly – and even corrected its own errors. I linked the project to GitHub for deployment to our hosting service, which it handled with ease.

The Not So Impressive

Initially, it struggled with React’s routing but quickly corrected itself. It also had cross-site scripting issues, but once I pointed these out, it fixed them and retested the site’s security. However, these were minor issues, easily resolved.

Project 2 – Creating a Simulation in R with Claude.ai

The Task

Alongside beekeeping, I lecture on bee genetics. One goal was to show students how to model quantitative genetics. This required a model of colony growth and breeding under different conditions in R – a more conceptual task than website building.

I chose Claude.ai for this project. An R library, SIMplyBee, already existed for modelling honey bee breeding. My plan was to have ChatGPT craft a suitable prompt, then let Claude generate the code. After multiple iterations, I had a prompt that seemed robust enough for Claude to produce a complete solution, over one thousand lines of R code.

The Impressive

Claude was excellent at producing code. I could change the logic – logistic growth, stochastic variation, environmental effects – by simply rephrasing prompts. Its ggplot2 outputs were visually appealing, labelled, and stylistically polished. After extensive tweaking, I ended up with a compelling simulation that modelled hybridisation in honey bees with colourful charts – perfect for lectures.

The Not So Impressive

Initially, the R code had a number of syntax and logic errors, and it referenced non-existent functions – classic AI hallucinations. When I asked for biologically meaningful complexity, inconsistencies multiplied: duplicated code blocks, misused parameters, and logical contradictions. Eventually, I had to clean and integrate the pieces myself.

Comparing the Two Experiences

AI performed beautifully when the task followed clear rules and patterns. Web design adheres to conventions, and the AI knew them well. However, it faltered when reasoning or domain knowledge was needed – areas where correctness depends on meaning, not syntax.

We often hear about ‘AI pair programming’ as if AI were an equal collaborator, but it is more like mentoring a hyper-fast intern: brilliant at pattern recall, occasionally reckless, and completely devoid of judgment.

Both projects revealed the same truth: AI can write convincing code, but it cannot yet understand problems. Lovable.dev mastered structure; Claude.ai mastered syntax. Neither mastered intent. That gap is where human developers are still essential – we provide judgment, context, and meaning, the qualities that transform code into understanding.

The Economics of AI Tooling

For organizations, the implications extend far beyond technical curiosity. AI-assisted development is a major shift in the economics of software engineering. Projects like those described here suggest that AI tools can significantly improve developer productivity, code consistency, and codebase maintainability. Routine scaffolding and code review can be automated, freeing skilled engineers to focus on architecture and quality, helping to reduce project timelines, improve onboarding, and make codebases more standardised and testable. However, realising those benefits requires strategic investment in process redesign, training for AI literacy, and clear accountability frameworks to ensure that human oversight is still central to quality and security.

About the Author