Reflections After 3 Years of Use
I've been intensively using generative AI since ChatGPT was released in November 2022. Software development quickly became one of the areas where I use it most. Repeated cycles of enthusiasm and disappointment have led me to a more mature understanding of its strengths and limitations. Here are some reflections.
The New Baseline for Development
I'm convinced that in the current state of the art, specialized code generation agents are a fundamental tool for software developers. Those who don't adopt them are creating a massive handicap compared to those of us who use them systematically.
Beware of Trivialization
I'm concerned about the trivialization happening when people without software engineering knowledge see how "with a simple prompt" they can build functional applications "without coding." While this is possible, there are several caveats to keep in mind.
Many platforms that let you generate applications with a prompt lock you into their environment (vendor lock-in) and produce applications with security, scalability, and maintainability issues unsuitable for production. On top of that, people without software experience lack the ability to write complete specifications, so the AI fills in the gaps with assumptions that aren't made explicit and often don't align with expectations. These applications can be useful as prototypes or for private use, but risky for broader deployment.
The Senior Developer's Superpower
On the flip side, software development agents can be a superpower in the hands of a senior developer. AI cannot produce production-grade software without proper guidance, but it can dramatically accelerate productivity.
This isn't about "vibe" coding, but about spec-driven development. That means producing detailed technical documentation that guides development agents to generate the required code.
What does it include? Functional specifications with measurable acceptance criteria, architectural patterns, code quality standards, testing, security, observability, among others.
The developer's role doesn't disappear; it evolves from writing code to designing the structural elements that enable quality software. As Faisal Haque puts it, this is the evolution from coder to builder.
I've been dedicating myself to exploring and experimenting with how to produce high-quality software using agents, and I'm preparing a series of posts to share the learnings I've gathered along the way.
In the next posts, I'll dig into the specifics: how models actually behave, what spec-driven development looks like in practice, and the techniques that make the difference. What's been your experience so far?
