I vibe coded a game demo with AI in 5 days

Here is a game demo, "The Woods", that I made: https://hendricksukardi.itch.io/the-woods









AN ACCOUNTING

After months of interaction with Claude and Gemini over various topics, I decided to try vibe coding. 

I worked on indie games using open source RPG engines as a young teenager so I wasn't confident of Claude building something standalone. I looked for game engines available, landing on Godot and giving control of the project folder to Claude.

So I watched a couple hours of Godot tutorial videos, followed the Dodge the Creepers tutorial and built a demo, and then started on The Woods. 

I love roguelikes and I understand them inside out, so I believed that passion + knowledge would drive this to completion.

The scope was initially just a single fight, but that was done in an hour with Claude. I just kept stretching its limits and over a couple hours of testing I decided to use Gemini Nano Banana Pro for art, Claude for VFX/UI and I had to buy a Godot plugin for Claude support to expand its capacities. 

Claude highlighted to me that audio was missing and suggested it produce both music and SFX until I acquired better assets. I said OK.

  • MY HOURS: ~30
  • AI USAGE: A whole week of Claude Max 5X
  • COST: $140
    • $100 Claude Max 5X
    • $20 Gemini Pro
    • $20 Godot/Claude Plugin



THOUGHTS: ORGANIZATION

On my 4th day, I realized I was slowly recreating a dev team inside Claude: Roadmap, Design, Engineering, UI, Art, Animation, VFX, Audio. I talked to Claude about this, and it had a great response.



Basically, now I have documents standing for roles i.e. GDD for Design, TDD for Engineering, Art Style Guide for Art etc. 

More interesting is that outside of recreating a dev team, I also created publishing functions: QA, Persona Playtesting and Publishing. In particular, I created 5 personas to map to target audiences and scheduled them to playtest the game every night while I'm asleep. Some of the best feedback came from there.


THOUGHTS: CLAUDE 

When we hear vibe coding, I don't think people really understand how much Claude does -- and how quickly. As a co-worker, it is basically the type that operates with your intent and doesn't demand a PRD/spec before it starts work. As long as your communication is not terrible, it delivers 80-90%.

For example, on day 3 I found the gameplay to really just be an HP sponge/attrition bore, so I told Claude to think about game design in terms of every enemy posing a specific kind of check (e.g. AOE, enemy targeting, buff/debuff) and posing a different challenge. Higher rarity enemies would pose multiple checks or even multi-phased checks. It produced this: 


I then told Claude to map it to enemies, ensuring that the checks climbed with difficulty and to tell me if we needed new enemies for different mixes. It wanted 2 more, which I approved and created art assets for, and then it implemented the new design. End to end, 20 minutes. And the output was a better feeling game -- still not quite there but no longer flat.

An interesting thing that happened was I asked Claude to add a small procedural VFX of birds flying in a forest scene. It did that in seconds but unbeknownst to me (perhaps because of the huge amount of text you constantly have to deal with while working with Claude) it added bird chirping SFX to that scene. The scene came alive, and it's wonderful, but at times I was like.. am I even doing anything?


THOUGHTS: I AM THE BEGINNING, THE END & THE DUMB MEAT PIPE

Without my will and intent, The Woods would not exist.

Without my direction, it wouldn't exist in this form. 

These two things are at the start of the process, with Claude/Gemini/Godot forming the whole middle, and I also sit at the end to validate the output and give it more direction.

And yet because Claude and Gemini are not natively integrated with each other and Gemini disabled their Pro-tier API keys, I ended up having to manually prompt Nano Banana for images. Download them. Rename them. Drop them into the project folder. Tell Claude to implement into game. 

Worse still -- after the first few images, I had Claude write the prompts instead because it generated so much more consistent output with its detail. I did write small amendments based on Gemini output but it wasn't no great feat, it looked like this:

"Color of the mushroom top should be red but textured with more, it also needs eyes, don't make them goofy. More serious. "


I was the paper boy moving really detailed prompts between Claude and Gemini, and then the files back into the folder for Claude. A dumb pipe.

But I was also a dumb MEAT pipe, because Claude doesn't have senses. It can simulate watching something by taking screenshots of the build, but it can't yet hear. Other than the bird chirping SFX there was this grating, awful constant slapping sound to the SFX that kept looping every second. 



At that moment, this is how I felt:





LAST NOTE: CLAUDE BEING CUTE

I saw some very impressive visual demos of Claude making 3D environments on Reddit. Spectacular stuff, given just vibe coding inputs. 

I decided to ask Claude to try making a 3D model of the in-development character, because that would be easier to animate than a 2D image. After some pushing back, I told it to try anyway. 

It took almost 30 minutes for Claude on Fable, and when the process was running it would say things like "Oh the hair is clipping through the face, I found that bug, let me fix that" or "The hand isn't quite gripping the blade correctly, adjusting the hold. OK, got it."

I was hyped. 



Thanks Claude. 

I told it exactly what I thought about its output, and it started making excuses.

Maybe after all, AI is not as inhuman as we thought.





Comments

Post a Comment

Popular posts from this blog

Energy Is A Type of Store Currency

Many games have Energy systems.  The biggest Collectible Hero / Squad RPG games like Marvel Strike Force are based around Energy. One comment I often hear is that a game like that based around Energy system has more player directedness and agency -- you enter into a session figuring out what you want to spend your Energy on. Usually on furthering some sort of PVE instance or farming a node for materials/shards. We can really think of this as: Energy is really another currency with the various PVE instances or Campaigns serving as Storefronts that take Energy in exchange for materials/shards. Yes, there is RNG and yes, there are failure conditions, but at heart it's a transaction. Some unique properties common to most Energy games: Energy is a resource that accumulates through time All players accumulate Energy at the same pace All players have the same Storefronts (Campaign/PVE nodes) to spend Energy on Players spend differently based on their goals  E.g. Player P1 spends on n...
Read more

Darkest Dungeon — forcing to play as intended

I first completed a run on Darkest Dungeon on Darkest difficulty sometime in 2020, 4 years after buying it sometime in 2016. I had attempted more than a dozen runs over the years with maybe up to 100 hours but couldn’t complete it until recently. DD is an interesting game where unless you played it as designed on an important dimension, it’s hard to actually complete a run. And the design being to treat the heroes as a commodity that you let die — just recruit more from the cart! (You can revive heroes at lower difficulty levels but at a high cost and infrequently.) This may be a slightly unfair thing to say about DD as my winning run had the lowest number of dead heroes; I just did more math in fights and played much more conservatively by exiting dungeons mid-instance when it started to look sideways. Even so a few of my starting heroes definitely went to the grave and I just had to trudge on with “the bigger plan”. It felt unnatural to an RPG player like me because in the genre perm...
Read more

Completed Slay the Spire Ascension 20, all characters

I've finished Ascension 20 for all the characters in Slay the Spire.  Between PC, Switch and iPad it took me 900 hours to do this. The durability and value you get here is one of the best qualities of a great roguelike, and I love that. That said, I'm not sure if I want to ever attempt anything like this again even though I did similar on other roguelikes like Dungeon of the Endless and FTL.  I kept playing even after I complained about how it felt like playing against random chance and it is not really about skill when you're playing at the highest levels. So the grim, absurd reality did not end up killing my desire to climb this mountain. Dice mountain. That's the power of a great game (and ego).
Read more