What Happened
Developer Carlos Menezes published a post titled "My sentiment on AI, March 2026" laying out where AI coding tools genuinely help and where they fall flat. His core argument: AI cannot shortcut the foundational work of defining architecture, understanding constraints, and building the mental model of a system. That part is still entirely human.
Where AI shines, he says, is in small, self-contained tasks - utilities, refactors, isolated bits of functionality. Once you have the architecture laid out and know what you need, AI can fill in the pieces fast.
He tested this by building Apple Watch apps with almost no Swift knowledge - a score tracker for 5-a-side football and a tennis service speed approximator. AI handled the unfamiliar language well because the scope was contained and the codebase was small.
His summary: "The work doesn't disappear, but the parts that used to take hours sometimes take minutes."
Why It Matters
This reflects a sentiment that is becoming the consensus among working developers in 2026. The initial hype cycle around AI replacing programmers has cooled into something more useful: a practical understanding of where these tools fit.
For anyone building AI-assisted workflows, Menezes confirms what daily users already know. AI coding tools are best at brownfield tasks with clear boundaries. They struggle with greenfield projects where the hard work is deciding what to build, not writing the code.
This matters for tool selection too. If you are picking between AI coding assistants, the differentiator is not raw code generation ability anymore. It is how well the tool understands existing project context and handles scoped tasks within a larger system.
Our Take
Menezes is right, and his framing is useful. The "AI as junior developer who needs clear instructions" mental model holds up well in practice. You still need to design the system, break down the work, and review the output.
The Apple Watch experiment is a good illustration. AI let him skip the "learn Swift basics" phase and jump straight to building. That is a real productivity win - not because AI replaced engineering judgment, but because it handled the syntax-level work in an unfamiliar language.
If you are using tools like Cursor, Claude Code, or Cody, you have probably already arrived at this same conclusion. The best workflow is human architecture decisions plus AI implementation of bounded tasks. Anyone still expecting AI to handle the full stack from design to deployment is going to stay disappointed.