“Go to Brazil” is a small experimental website I built to explore how far it is possible to create a digital product using only artificial intelligence tools. The result can be seen publicly here:
https://facundouferer.github.io/gotobrazil/en/
Beyond the site itself, the real interest of the experiment was to observe how a fully AI-mediated development workflow behaves and, above all, to identify the point where current tools still leave a gap they cannot completely cover.
The premise of the experiment
The idea was simple but strict: everything had to be generated or assisted by artificial intelligence.
The human role was reduced to:
- defining the general project goal
- coordinating the tools
- executing the final deployment
No part of the site — not the design, planning, code, or images — was supposed to be created manually without AI assistance.
Step 1: Initial visual design with Stitch
The starting point was the visual design of the site. For this I used Google Stitch, an interface generation tool powered by artificial intelligence.
The AI produced the first graphic proposals for the site: homepage structure, sections, visual style, and content organization. At this stage, elements such as the following were defined:
- overall site layout
- section hierarchy
- tourism-oriented visual style
- primary content blocks
This first step made it possible to obtain a visual base quickly, something that would traditionally require the work of a UX/UI designer.
However, as often happens with automatically generated interfaces, the result was an initial approximation rather than a fully polished final design.
Step 2: Project planning with Claude
Once the visual concept was defined, the next step was to structure the whole project.
For that I used Claude, which generated a detailed plan for the site:
- page structure
- expected content for each section
- project architecture
- tasks required for development
- a complete list of implementation issues
In other words, Claude acted as a kind of product manager and software architect, transforming the initial idea into a structured work plan.
This stage was particularly interesting because it turned a diffuse concept into a concrete development roadmap, something that normally requires experience in software project management.
Step 3: Issue generation with Codex
Based on the planning generated by Claude, I used Codex to transform that plan into concrete development issues.
Each issue represented a specific task:
- creating interface components
- implementing site sections
- integrating images
- adjusting styles
- preparing the content
The goal was to simulate the workflow of a real development team, where work is organized into small, clearly defined tasks.
This way, artificial intelligence did not only help generate code, but also helped structure the work required to produce it.
Step 4: Image generation with Nano Banana
The visual material of the site was generated using Nano Banana, an artificial-intelligence-based image generator.
All illustrations and images used on the site were produced from prompts related to:
- Brazilian landscapes
- tourism
- beaches and nature
- tropical aesthetics
This made it possible to complete the graphic material quickly without using stock image banks or manual design.
Step 5: Full development with OpenCode
The implementation of the site was carried out using OpenCode, working with several open and free code models.
The goal was for all code to be generated automatically from the previously defined specifications and issues.
The process consisted of:
- taking each generated issue
- asking the agent to implement the functionality
- reviewing the result
- adjusting the instructions when necessary
This process reproduces an interesting dynamic: instead of writing code directly, the developer describes what is needed and the agent implements it.
In that sense, the human role shifts from programmer to director of the software generation process.
Step 6: Deployment on GitHub Pages
Once development was completed, the site was deployed using GitHub Pages, which made it possible to publish it quickly as a static site accessible from the web.
This was practically the only fully manual step in the process: executing the final deployment.
The result
The experiment showed that today it is possible to build a complete website using artificial intelligence tools exclusively for:
- design
- planning
- task generation
- image production
- code writing
In terms of productivity, the process was extremely fast compared with traditional development.
However, the real objective of the experiment was not to prove whether AI can program — something we already know it can do — but to detect where the real limitations appear.
The gap artificial intelligence still does not cover
The main conclusion of the experiment is that code generation is not the main problem.
Current tools can produce functional code quite easily.
The real challenge appears elsewhere: the global coherence of the project.
During development, a difficulty kept appearing that could be described as the major current gap in AI-assisted programming:
the lack of deep understanding of the complete system.
Each tool can solve individual tasks very well:
- designing a screen
- generating a function
- creating an image
- writing a component
But no tool still maintains a persistent, integral view of the whole project.
This creates several problems:
- inconsistencies between components
- contradictory architectural decisions
- loss of context across tasks
- difficulty maintaining aesthetic or structural coherence
In other words, artificial intelligence is excellent at solving local problems, but it still struggles to manage the systemic complexity of a complete project.
The new role of the developer
This experiment suggests that the human role in AI-assisted development does not disappear, but instead shifts toward another kind of task.
The developer becomes someone who:
- maintains the coherence of the system
- defines the problem clearly
- supervises architectural decisions
- corrects deviations in the automated process
They are less a programmer in the traditional sense and more a technical director of generative systems.
An unexpected conclusion
The “Go to Brazil” experiment shows that AI-assisted programming is already a practical reality.
A complete website can be built with almost no manual coding.
But it also reveals something more important: the biggest challenge is no longer generating code, but maintaining the structural intelligence of the project.
Until tools can understand and sustain that global vision autonomously, there will continue to be a fundamental space for human judgment.
And that is probably, for now, the real job of the programmer in the age of artificial intelligence.