I Asked Grok 4.6 to Build me an App..
A tool for writers who hate AI prose
Most writers don't want AI to write their stories. They want help with the boring parts of worldbuilding, like making a map where the rivers actually make sense. Vish, a data science intern at Super Bash, built a web app in Cursor with Grok 4.6 that tackles exactly that. You doodle continents and islands on a canvas, adjust brush size and planet radius, then hit critique.
The tool tears the map apart. It runs about 300 tests against real geographical data and returns a score. Too many small islands? Continents full of holes? A river flowing uphill? It flags all of it and then offers a fix that closes the holes and smooths the landmasses into something that could pass for a real planet. From there you can render the map in 3D, check elevation, biomes, moisture, and temperature, or even view it in winter. You can also place settlements and get auto-generated descriptions like farmland, pastoral town, or mine. The whole thing runs on a MacBook Air.
Where Grok 4.6 fell short
The build moved fast, but the first version was surface-level believable. Vish called it 'a doodle, not a planet.' The back end looked fine on the surface, but the geography didn't hold up. One version placed cold areas right next to hot ones. Another put mountains where tectonic plates wouldn't collide. Grok built pieces separately and didn't wire them together coherently.
AI tools are great at smoke tests. They make buttons click and pages load. They handle CSS and frontend patterns because they've seen a million websites. But this world builder was niche. There wasn't a training set full of fantasy geography simulators to copy. So Grok picked the simplest architecture for itself, not the right architecture for the job. Vish had to go to a real person who knew geography, run separate tests on math and rendering, pull templates and skills from GitHub, and wire everything together himself. He also had to stop working with subagents in parallel when one feature needed the output of another. Plates come before mountains. Small changes ripple.
What this means for your own projects
The practical takeaway: Grok 4.6 is fast and can produce a lot of features quickly, but it needs supervision. It's not AGI. For a niche tool, you supply the domain patterns, the order of operations, and the reality checks. The AI speeds up the build. It doesn't replace the thinking.
The next steps for Vish are visual polish, since the older versions looked better, and then layering on more complexity like trade routes. The core proposition stays the same: let doodlers make a mess, then help them turn it into a world that makes sense.