← sources

I tried building GTA Manila before GTA 6, using Claude

source: reddit (reddit:ClaudeAI)published 2026-11-18 07:37ingested 2026-11-19 02:36

retained text

This was a fun experiment I started 2 weeks ago. I was tinkering with AI-assisted workflows in Unreal Engine. I thought it would be really funny if someone tried to make GTA Manila before GTA 6, right before the meme ends.

The project is unfinished. I only have a single $100 Claude Max subscription and a $20 ChatGPT subscription. I did not use Tripo 3D, Meshy, or Hunyuan. I knew making an open-world game with realistic graphics set in the entirety of Manila within 2 weeks was unrealistic. I settled with the following assumptions and compromises:

1. Single-player only

2. Used free assets from Unreal Fab, and assets that I already owned. With limited token budget, it was impractical to create all the 3D models and animations from scratch using AI. I found a bunch of useful shaders, character animation, and vegetation assets that are worth it.

3. For custom 3D models, limited review loops to only 1-2. Claude as the 3D modeler while GPT-6.1 Sol Xhigh as the critique. This did not produce the results I wanted, but hey, I gotta work with what I have. Too many loops would burn through my tokens fast.

4. Only selected iconic locations of Manila will have high details while everything else is low-poly.

Moreover, I was working on other projects besides this one. Not all of the Claude Max subscription was used here. In the end, I was not able to implement the features I wanted. There are still many missing systems.

**Lessons learned:**

1. Having a good prompt and excellent reference images can carry the quality of the output by a lot.

Probably sounds obvious to most of you, but to me the biggest mistake I did were the lousy skills and references I was using during the early stages. I made the mistake of adding too many rules which prevent Claude from trying different approaches. Use /doctor and double check that your setup follows the Claude Opus and Fable prompting guide. Regarding reference images, the reasoning is in the next item below.

2. Art Direction > Realism. 

This applies especially to lighting, post-processing, and visual effects. A wise 3D artist once told me, "Great lighting and post-processing can make mediocre 3D models look good". It's certainly possible to get great 3D models by giving Opus a lot of reference images and have it undergo multiple review loops with itself or GPT-6.1 Sol. However, when putting all the objects together and assembling the scene in Unreal, my initial results looked bad because Claude was taking inspiration from real-world location lighting. I recommend gathering references of lighting setups in modern video games, film, or edited photographs to deliver the mood and atmosphere you're trying to achieve.

3. Fable 5.2 is INSANE, but deym is it a token furnace. 

I know Fable can one-shot entire games that look amazing. However, with the limited amount of Fable tokens I have, I used it primarily on lighting, post-processing, cutscenes, and character animations.

4. Start building your personal 3D asset library.

Lots of free high-quality assets out there. Create your own asset browser tool and have your agents scan through it recommend you assets before attempting to create something.

5. Texture shimmering, light and shadow flickering, z-fighting, Lumen splotches.

These things can sometimes go under the radar without getting fixed because Claude only analyzes a couple of frames not in a sequence.

**Music**: “Light it Up Boy” by Michael Ramir C.

document fetched 2026-11-19 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xz9anh/i_tried_building_gta_manila_before_gta_6_using