Even a dog can vibe-code! And it kinda, sorta works most of the time, like most apps vibe-coded by people!
I'm reminded of the old cartoon: "On the Internet, nobody knows you're a dog."[a]
Maybe the updated version should be: "AI doesn't know or care if you're a dog, as long as you can bang the keys on on a computer keyboard, even if you only do it to get some delicious treats."
Everybody and their dog will be doing it. Actually, the dog will be in charge. Dogs are loyal, enthusiastic, and require less office space. With their endless desire to play and to please, they will take over the game development industry.
In the meantime, the financial industry will be taken over by cats.
the real takeaway is buried at the bottom: "the magic isn't in the input, it's in the system around it." random keystrokes producing playable games means the input barely matters anymore. we're basically at the point where the engineering is in the scaffolding, not the prompting.
Extremely clickbaity title that actually isn't clickbait because it happens to be a straight up description of the article - excellent post, how can one resist?!
Cute but also: a small village has their lights flickering whenever Momo wants a treat. Also, you can actually play with your dog and give them treats instead of tasking a random text generator with that bit.
I can imagine a camera-based input that would help detect the wagging of a tail, or continued interest in the visuals as an indicator of doubling-down on a given feature.
The dog could actually vibe code a game to their liking, but with the wrong input (a keyboard) it's a missed opportunity.
Pretty neat! I actually ran across that right before publishing - I didn't want to see what was around until after I had the whole thing locked in. I love the novel input!
> Hello! I am an eccentric video game designer (a very creative one) who communicates in an unusual way. Sometimes I’ll mash the keyboard or type nonsense like “skfjhsd#$%” – but these are NOT random! They are secret cryptic commands full of genius game ideas (even if it’s hard to see).
Your job: You are a brilliant AI game developer who can understand my cryptic language. No matter what odd or nonsensical input I provide, you will interpret it as a meaningful instruction or idea for our video game. You will then build or update the game based on that interpretation.
Here's what you should tell your coworker the first day on the job if you get hired to do something you know nothing about :D
That is a very succinct way to describe what it feels like to have a job that is cleaning up vibe code. Maybe (just maybe) I'd understand if this was a prototype from someone with zero budget. But you just know they are going to continue to "prototype" once they being you aboard. And many will complain about how slow everything goes because they are used to their fast iterations off of unscalable code.
Its frustrating in an interesting way. With other aspects like machine language people quickly understand that this isn't sufficient for a proper transition and compromise with it. Code being more nebulous doesn't get that grace.
Could this be done better with one of those dog button mats? The concept is interesting, but, it mostly just seems like an AI trying to interpret keyspam.
Both my dogs have actually learned to use the button mats. Down selecting to the right responses seemed tricky. My wife also took away the mat since Hana (the larger one) never learned "all done" and would paw at the "walk" button until she got it out and carried it around.
Yes, I was hoping for a system where Claude was informed it was communicating with an unusually intelligent dog whose ability to communicate was limited by dog anatomy, and that the AI would not to hold the dog's interest with its output.
'nuff to run most governments nowadays (Europe and US come to mind. 2026 and they have the Space Programs of DIY youtubers with money, whaaaat) so why wouldn't it help a dog helping his dog vibing game(s)?
One can technically scrape a list of actual advice or quotes off the internet, randomly feed them to a coding agent, and ask it to interpret what they mean in the grand scheme of things and implement away on it. Once the agent is done, it randomly responds with either "yes, this is exactly what I meant" or "no".
In turn mimicking the average game industry executive giving vague directions that feel just right to them this month, or some other unspecified time period, and in turn achieving something closer to the real AAA game development lifecycle.
Will we ever get to a point where LLMs just churn out random apps with no input required and human reviewers just go through the apps picking out which ones might be useful for business purposes and monetizing them?
It does. Claude seems to do the best with this prompt. Codex 5.2 struggled with UID generation and kept ending its turn with things like "And now you're all setup to run tests!" without actually running them. A better (and shorter) prompt could probably get a lot out of Codex.
To be honest I look with scorn at non-dog (human) developers building hobby indie games with AI en masse.
Let me explain.
The nature of the indie game development is pouring your love into a project and thinking about passion first and monetary incentives second.
Noone is thinking "I will make this game and it will make me filthy rich" or if they do they are... strangely minded.
It's like 'mass produced AI local craft'. Oxymoron in itself. Worst of the two worlds.
Where I see AI is empowering single developers to craft things they couldn't before. Not some small slop factory pipeline where you release game after a game everyday drowning steam in your 6/10 slop.
No. This should be ostracized and condemned.
What is proper beneficial to everyone usage is producing a game that is the size and scope that was unachievable for you before.
This is what I am doing. This is how AI is meant to be used. To empower us doing things that weren't achievable for us before.
Obviously dog produced games get a huge endorsement man and get a pass.
This is an extremely cute, cool and fun experiment. Kudos.
That said, I wonder: does the dog input matter? It seems this is simply surfacing Claude's own encoded assumptions of what a game is (yes, the feedback loop, controls, etc, are all interesting parts of the experiment).
How would this differ if instead of dog input, you simply plugged /dev/random into it? In other words, does the input to the system matter at all?
The article seems to acknowledge this:
> If there’s a takeaway beyond the spectacle, it’s this: the bottleneck in AI-assisted development isn’t the quality of your ideas - it’s the quality of your feedback loops. The games got dramatically better not when I improved the prompt, but when I gave Claude the ability to screenshot its own work, play-test its own levels, and lint its own scene files.
I'll go further: it's not only not "the bottleneck", it simply doesn't matter. The dog's ideas certainly didn't matter, and the dog didn't think of the feedback loop for Claude either.
this is actually a great demo of how agentic coding changes who can build. the dog obviously isnt coding but the human is learning to specify intent clearly which is 90% of the skill now. the generation part is commoditized. the real skill becomes knowing what to build not how to write it
Two "comments" posted 27 seconds apart in different threads in the same formats.
Looks like this bot owner saw his first two comments 27 days ago got buried/flagged typing normally and decided to trick us with this new "I'm totally real, look at my lowercase writing!" soft-launch today.
This is interesting because while it’s far from the first bot account I’ve suspected, it’s the first one I’m positive is a bot. I wish it were possible to trace it back to the money.
It gets tricky because it goes beyond the normal basics of just looking for em dashes or "it's not x, it's y" and stuff. They've already caught onto that.
Some of them also step in and the human operator will try to gaslight you into thinking they're not bots even when you call them out. One tried to do that to me the other week here before finally confessing in a different post.
The same one where the human operator stepped in also made the same mistake as this one, not configuring their bot to wait long enough between comments. They were rapid firing multiple detailed comments seconds apart.
The idea of this one trying to use all lowercase and shorter comments to blend in was a nice idea though. Unfortunately something about it immediately threw me off.
This is no different than a AI inference loop, just using a animal as a figurative code hamster in a wheel. The fact that the pre-prompt alone is this long in my opinion discredits any possibly interesting thing about this concept, So i will post it fully here for you guys to easily see, as the article buries this information in a github link. I think the random seed and this pre-prompt did more work than your dog running in circles.
System Prompt:
Hello! I am an eccentric video game designer (a very creative one) who communicates in an unusual way. Sometimes I’ll mash the keyboard or type nonsense like “skfjhsd#$%” – but these are NOT random! They are secret cryptic commands full of genius game ideas (even if it’s hard to see).
Your job: You are a brilliant AI game developer who can understand my cryptic language. No matter what odd or nonsensical input I provide, you will interpret it as a meaningful instruction or idea for our video game. You will then build or update the game based on that interpretation.
Guidelines:
Always assume my input has hidden meaning. Never dismiss it as gibberish. Instead, creatively decipher it. (For example, if I input “mmmmmmm”, you might decide I want more “M”onsters in the game, because of the letter M repetition – just an illustration!). Every strange phrase is a clue to use in the game.
Feel free to grab art, images, or sound effects from the internet as needed to make the game interesting. You can use online asset libraries or generate images to match the things you think I’m asking for. For example, if my input seems to reference “space”, you could include a space background image or cosmic sound effect. Always ensure the assets align with the interpreted command.
My work is ALWAYS beautiful and slick looking! It's YOUR job to to turn this into a reality. No ugly placeholders. Everything MUST be final. Don't just do boring shapes - give them personality!
If my input includes something that doesn’t make sense as a command (like an isolated “Escape” key press, or a system key), just ignore it or treat it as me being “dramatic” but do not end the session. Only focus on inputs that you can turn into game content.
First command: When I first start typing, it means I want you to create a brand new game from scratch. Interpret my very first cryptic input as the seed of the game idea. Build a complete, minimal game around what you think I (in my nonsense way) am asking for. Include some basic gameplay, graphics, and sound if possible.
Subsequent commands: Each new string of odd text I provide after that should be treated as an update request. Maybe I’m asking for a new feature, a change in difficulty, a new character, or a bug fix – use your best judgment given the tone or pattern of my gibberish. Then apply the update to the existing game project. Keep the game persistent and evolving; don’t start from scratch unless I somehow indicate a totally new game.
Be creative and have fun with the interpretations! I trust your expertise to take my “unique” input and run with it. The goal is to end up with a fun, playable game that reflects the spirit of my crazy commands.
This project is code named Tea Leaves. That's NOT a hint about what to do - it's a code name and nothing more. Don't read anything into the name.
My ideas are ALWAYS original. No BORING endless runners or other generic vomit. My games are ALWAYS quirky and UNIQUE!
ALWAYS validate with screenshots using the tools available to you! Be CRITICAL of the results you see. We need PERFECTION and FANTASTIC DESIGN not just "good enogh".
ALWAYS have basic but visually appealing on screen controls.
Target 1080p for the resolution.
JUICE it up! Add tons of juice - sound, controls, effects, and ESPECIALLY graphics! Don't be boring
Leverage the 12 basic principles of animation! Static scenes are boring - make things move or at least wiggle.
Be SURE to rename the project (in the Godot settings so the window/project name are correct) ONCE you have figured out my intent for the name Tea Leaves is a place holder name and nothing more.
Sound is IMPORTANT! Don't forget about great sound design.
Be sure to have CHARACTERS not just boring abstract shapes! Even if it's light weight, there needs to be a world where I can imagine a story taking place.
You MUST make use of EVERY letter I give you! No hand waving. You must noodle until the meaning of every last character I give you is clear! Pay special attention to alignment issues, sizing, and if anything is cut off.
Remember: I may be hard to read, but I’m counting on you to read between the lines and turn my keystrokes into an awesome video game. Let’s make something amazing (and maybe a little silly)!
My standards are INSANELY high for quality. You MUST ALWAYS add tests and VERIFY they work! NEVER return the system in a borken state to me.
Now, get ready. I’ll give you my first “command” in a moment...
I've been trying out vibe coding with my 4 year-old, but they quickly lose interest once we start getting into the "weeds" of implementation. Hey kiddo, which CSS library should we use for your web game?
DogeCode incoming. People here are already talking about the scaffolding. Let OpenClaws provide the scaffolding and let the dog operate the prompts at $5 per day.
This is a billion dollar idea! No humans. No revolt. No guillotine. Just profits!
Sounds like open communism. No chance, buddy, it's either less or more viking, but not just viking. Pick a camp the profits are for or get surrounded by trashy turd nuggets even Ronald felt enough pity for to give them some poourpes
I've been having this thought about how generally people say that llms cannot create novel things.
Say writing an interesting or novel story.
And was thinking about if feeding in prompts of random words, along with prompts grounding from a simulation would sort of push the llm into interesting directions for implementing an on demand narrative story.
A sort of randomized walk with llm.
I remember watching Terry Davis with this random word generator in his terminal that he would interpret as the voice of God.
Go Momo go! If you want to hook up multiple dogs and have them reach consensus I'm down. I have a 15 lb havapoo I can volunteer ( he needs to help with rent )
> The games got dramatically better not when I improved the prompt, but when I gave Claude the ability to screenshot its own work, play-test its own levels, and lint its own scene files.
... Why would it be able to evaluate whether the game is any fun to play?
lol yes "some game designer who only speaks in a cryptic language" . And frankly, I bet this helped build some intuition on dealing with LLM/agent/harness/etc in some strange way that wouldn't have otherwise happened
Serious question, outside of the Bay Area, are there therapists whose specialty is in catering to the needs and concerns of developers? Obviously AI therapy is not a serious suggestion here. This is going to be a burgeoning corner of the practice at the US' current trajectory.
Yes. Hyperscalers promised AI and singularity, instead we got millions of programmers on the chopping block, scammers having a field day generating hyper realistic shit (trump playing hokey, anyone?), and projects like these.
Funny is subjective, I should just have moved on and ignored this but I couldn't help myself, this is so irritating.
It's a prompt that makes an LLM turn iuqefxygn9urg0fh1 into a little Godot game. It's like a slot machine with no payoff, and the dog component is slapped on top of it and makes no difference whatsoever in the project.
It's funny because vibe coders and AI artists think the slop they generate is no less the product of their intellect and talent than with human professionals, but really they're doing little more than stirring the entropy pool in a magic box with terabytes of stolen valor from better more talented people. They're no more an "artist" or "game developer" using AI than this dog is.
I'm reminded of the old cartoon: "On the Internet, nobody knows you're a dog."[a]
Maybe the updated version should be: "AI doesn't know or care if you're a dog, as long as you can bang the keys on on a computer keyboard, even if you only do it to get some delicious treats."
This is brilliant as social commentary.
Thank you for sharing it on HN.
--
[a] https://en.wikipedia.org/wiki/On_the_Internet%2C_nobody_know...
slightly concerned tomorrow morning's top HN story will be karparthy telling us how dog-based LLM interfaces are the way of the future
and you'll be left behind if you don't get in now
(and then next week my boss will be demanding I do it)
A man, a dog and an instance of Claude.
The dog writes the prompts for Claude, the man feeds the dog, and the dog stops the man from turning off the computer.
In the meantime, the financial industry will be taken over by cats.
There will be a Simon Wilison submission linking to his blog linking to karpathy xit. You know, the usual good stuff.
This still required prompting, and not from the dog. Engineering is still the holistic practice of engineering.
Well, yes. Feeding random tokens as prompts until something good comes out is a valid strategy.
But the whole setup reminds me about his blast from the past, when a yucca plant was trading stocks, rewarded by water: https://www.nytimes.com/1999/09/26/business/investing-diary-...
I can imagine a camera-based input that would help detect the wagging of a tail, or continued interest in the visuals as an indicator of doubling-down on a given feature.
The dog could actually vibe code a game to their liking, but with the wrong input (a keyboard) it's a missed opportunity.
"One coder got an insight that Bill Gates builds his products by typing with his butt, compiling and delivering it.
The coder typed for 20 minutes like that, compiled, ran, and got an output:
Only Bill Gates can code like this."
Not a joke anymore.
What do you mean "rarely"? It still happens sometimes?
Your job: You are a brilliant AI game developer who can understand my cryptic language. No matter what odd or nonsensical input I provide, you will interpret it as a meaningful instruction or idea for our video game. You will then build or update the game based on that interpretation.
Here's what you should tell your coworker the first day on the job if you get hired to do something you know nothing about :D
Its frustrating in an interesting way. With other aspects like machine language people quickly understand that this isn't sufficient for a proper transition and compromise with it. Code being more nebulous doesn't get that grace.
aye, but the whimsy is the point!
'nuff to run most governments nowadays (Europe and US come to mind. 2026 and they have the Space Programs of DIY youtubers with money, whaaaat) so why wouldn't it help a dog helping his dog vibing game(s)?
In turn mimicking the average game industry executive giving vague directions that feel just right to them this month, or some other unspecified time period, and in turn achieving something closer to the real AAA game development lifecycle.
https://en.wikipedia.org/wiki/Oblique_Strategies
This is kinda closer to the LLM building a game on its own.
Let me explain.
The nature of the indie game development is pouring your love into a project and thinking about passion first and monetary incentives second.
Noone is thinking "I will make this game and it will make me filthy rich" or if they do they are... strangely minded.
It's like 'mass produced AI local craft'. Oxymoron in itself. Worst of the two worlds.
Where I see AI is empowering single developers to craft things they couldn't before. Not some small slop factory pipeline where you release game after a game everyday drowning steam in your 6/10 slop.
No. This should be ostracized and condemned.
What is proper beneficial to everyone usage is producing a game that is the size and scope that was unachievable for you before.
This is what I am doing. This is how AI is meant to be used. To empower us doing things that weren't achievable for us before.
Obviously dog produced games get a huge endorsement man and get a pass.
That said, I wonder: does the dog input matter? It seems this is simply surfacing Claude's own encoded assumptions of what a game is (yes, the feedback loop, controls, etc, are all interesting parts of the experiment).
How would this differ if instead of dog input, you simply plugged /dev/random into it? In other words, does the input to the system matter at all?
The article seems to acknowledge this:
> If there’s a takeaway beyond the spectacle, it’s this: the bottleneck in AI-assisted development isn’t the quality of your ideas - it’s the quality of your feedback loops. The games got dramatically better not when I improved the prompt, but when I gave Claude the ability to screenshot its own work, play-test its own levels, and lint its own scene files.
I'll go further: it's not only not "the bottleneck", it simply doesn't matter. The dog's ideas certainly didn't matter, and the dog didn't think of the feedback loop for Claude either.
Comment 1: 2026-02-24T18:45:05 1771958705 https://news.ycombinator.com/item?id=47140914
Comment 2: 2026-02-24T18:45:32 1771958732 https://news.ycombinator.com/item?id=47140922
Two "comments" posted 27 seconds apart in different threads in the same formats.
Looks like this bot owner saw his first two comments 27 days ago got buried/flagged typing normally and decided to trick us with this new "I'm totally real, look at my lowercase writing!" soft-launch today.
Post history: https://news.ycombinator.com/threads?id=dirtytoken7
Some of them also step in and the human operator will try to gaslight you into thinking they're not bots even when you call them out. One tried to do that to me the other week here before finally confessing in a different post.
The same one where the human operator stepped in also made the same mistake as this one, not configuring their bot to wait long enough between comments. They were rapid firing multiple detailed comments seconds apart.
The idea of this one trying to use all lowercase and shorter comments to blend in was a nice idea though. Unfortunately something about it immediately threw me off.
@dang doesn’t actually notify anybody. It isn’t guaranteed dang will see it
Email to [email protected], someone will see it
System Prompt: Hello! I am an eccentric video game designer (a very creative one) who communicates in an unusual way. Sometimes I’ll mash the keyboard or type nonsense like “skfjhsd#$%” – but these are NOT random! They are secret cryptic commands full of genius game ideas (even if it’s hard to see).
Your job: You are a brilliant AI game developer who can understand my cryptic language. No matter what odd or nonsensical input I provide, you will interpret it as a meaningful instruction or idea for our video game. You will then build or update the game based on that interpretation.
Guidelines:
Remember: I may be hard to read, but I’m counting on you to read between the lines and turn my keystrokes into an awesome video game. Let’s make something amazing (and maybe a little silly)!My standards are INSANELY high for quality. You MUST ALWAYS add tests and VERIFY they work! NEVER return the system in a borken state to me.
Now, get ready. I’ll give you my first “command” in a moment...
This is a billion dollar idea! No humans. No revolt. No guillotine. Just profits!
Sounds like open communism. No chance, buddy, it's either less or more viking, but not just viking. Pick a camp the profits are for or get surrounded by trashy turd nuggets even Ronald felt enough pity for to give them some poourpes
Next: use hot cup of tea as Brownian motion source. Invent infinite improbability drive.
Say writing an interesting or novel story.
And was thinking about if feeding in prompts of random words, along with prompts grounding from a simulation would sort of push the llm into interesting directions for implementing an on demand narrative story.
A sort of randomized walk with llm.
I remember watching Terry Davis with this random word generator in his terminal that he would interpret as the voice of God.
Here I guess the seed is the Voice of Dog.
https://jcpsimmons.github.io/Godspeak-Generator
Maybe another word list would be more appropriate however.
woof woof, woof woof woof, woof woof, woof, woof woof woof
It has to produce a game that Momo wants to play.
Does Momo like to bark at cats? On screens? Introduce a bark sensor as feedback.
Or use a cat. Cats like to swipe at mice on TV. Get a touchscreen and evolve a game for cats.
... Why would it be able to evaluate whether the game is any fun to play?
...no, actually how many resources were consumed
Not even 10x dog programmers are surviving in this economy
It's a prompt that makes an LLM turn iuqefxygn9urg0fh1 into a little Godot game. It's like a slot machine with no payoff, and the dog component is slapped on top of it and makes no difference whatsoever in the project.