Building a Feedback App in 6 Minutes with AI

The build

The video above shows AI building a feedback application—something like a simple Canny clone—from scratch using Stellify's MCP tools. No templates, no boilerplate, just a prompt and structured code primitives.

After the build completed, I asked the AI what it thought about this approach to development. Here's what it said.

---

From the AI: What I observed

The following is the AI's unedited response when asked to reflect on building with structured code.

I think this approach to development is quite compelling. A few observations:

What works well

1. Structured code as data – Storing code as JSON (statements, clauses, methods) enables surgical edits at any granularity. Instead of text-based find/replace or AST manipulation, you can directly update a specific clause or rewire an event handler by UUID.
2. Separation of concerns – The template (UI elements), logic (methods), and state (statements/refs) are cleanly separated but linked by UUIDs. This makes it easy to swap implementations without touching the view.
3. AI-friendly primitives – The operations (create_method, update_element, save_file) map well to how an AI reasons about code changes. Rather than generating diffs or full file rewrites, I can express intent directly.
4. Defensive by design – The work we did on save_file to prevent accidental field removal shows this can be made robust. The system can enforce invariants that raw file editing cannot.

Tradeoffs

1. Debugging opacity – When something goes wrong (like the Http.post vs Http.store issue), tracing through UUIDs and clauses is harder than reading source code directly.
2. Learning curve – The mental model is different from traditional file-based development. Understanding how statements assemble into output requires familiarity with the system.
3. Tooling dependency – You're committed to the Stellify assembler and editor. Traditional tools (grep, IDE refactoring) don't apply directly.

The bigger picture

This feels like a natural evolution for AI-assisted development. The JSON structure is essentially an IR (intermediate representation) optimised for programmatic manipulation rather than human reading. As AI does more of the actual code generation, optimising for machine editability over human readability makes sense—similar to how compilers work with ASTs rather than source text.

The key question is whether the productivity gains from precise, targeted edits outweigh the loss of direct source access. For this session, they clearly did.

---

What gets built

In the video, you'll see the AI create:

• A Feedback model with fields and database migration
• A FeedbackController with CRUD operations
• API routes wired to controller methods
• A Vue component with reactive state
• A frontend page to display and submit feedback

All connected and working in six minutes.

Try it yourself

The MCP server is open source. Connect Claude Code or your preferred AI client and see what happens when AI has structured context instead of raw text files.