The Gate is a 2D action roguelite that begins when a gate opens in Seokchon Lake, in Songpa, the Seoul neighborhood where I live. I am making it on my own with AI, from design and art to development and operations, and the game now running is v2, with its story and systems rebuilt from scratch.
Play The GateOn the night before the first anniversary of his grandmother's death, a gate opens in Seokchon Lake, and from inside comes the song she used to sing. Following it, the protagonist, Hangyeol, is pulled into a summer long past, where a grim reaper in a traditional gat hat is waiting. There, souls trapped in masks are causing floods and wars. Players fight masked enemies, borrow the power of sleeping guardian spirits to break the bosses' masks, and send the freed souls to the people they miss, closing eight gates from Seoul through Korea's rivers and seas to cities overseas.
Getting here took two years. I nearly gave up twice along the way, and once I threw out the whole story.
After 30 years of playing games, I decided to make one
I have loved games since I was a child. After more than 30 years of playing them, I knew what I liked, and as a product designer, thinking about what people choose and why they come back had become my job. As the technology moved fast, that naturally turned into wanting to make a game that carries a story of my own.
My first attempt was in 2024, with the Godot game engine and Grok. Even getting a very simple game running took an enormous amount of time. There was not much I could do with AI back then, and I stopped.
In 2025, I tried again with Figma Make. I got as far as a convincing mini-game, but a game with my own world and enough depth to keep people playing still felt out of reach. I stopped a second time.
In 2026, seeing what people were making with Claude's Fable made me think this time might be different. I worked through a design document in conversation with Claude and built the first prototype.

The first time I played it, I felt for the first time that this could actually work. I could attack, dodge, and parry an enemy's attack. The boss I had asked for appeared, and defeating it led to a reward. The graphics were rough, but I was playing through combat that had only existed in my head.

That first version was about gates opening in modern cities around the world, with a protagonist defeating regional bosses to find out why. I showed it to friends, kept fixing it, and released it. The real problems started after that.
The first wall: the same game felt different to everyone
I asked friends to play and watched where they got stuck. A friend new to this kind of game could not beat the first boss, while another, used to roguelikes, kept topping the leaderboard. The difficulty I had tuned for myself was nobody else's.

So Novice mode gives more protection after a hit and adjusts HP and attack power in the player's favor. Higher difficulties raise enemy damage, speed, and projectile frequency, and remove the brief invulnerability after being hit.
Balance is never set once and done, so I also built an admin where I could change those values and deploy them myself. I wanted to fix what I felt during a run straight away, without asking an agent every time.
The second wall: people liked the combat, but nobody remembered the story
People often said the combat was fun. Almost no one remembered the story. "A gate opened in some city" is a premise anyone could come up with, and ask AI for a world and you quickly get convincing fantasy. That was exactly why I could not see what made it my game.
In the end, I decided to rewrite the story from the beginning, starting from the place I know best. I live in Songpa, and I pass Seokchon Lake almost every day. Fewer people than you might think know that the lake used to be the Songpa River. When the Jamsil area was developed, one channel of the Han River was filled with earth, and the water left behind became today's lake. There was once a ferry crossing on its banks, too, Songpa Naru, where people and goods moved back and forth.
The memory of a buried river wakes up wearing a mask. That line became the start of v2. The first boss is the Songpa River itself, and the place you return to after each run is a ferry landing. I changed the protagonist, too, from a hero saving the world to a grandchild following his grandmother's song. Defeating an enemy now also means letting someone go.

Folklore became rules, not decoration
Once the story changed, the rest of the game had to follow. Folklore pasted on as a backdrop shows quickly, so I built it into how places were chosen and into the combat rules themselves.
Instead of picking famous cities, I chose places where water had been buried or forgotten, like the Songpa River. At Cheonggyecheon, a stream buried under a road for decades, a Han River water ghost rises. At Uldolmok, where the current has stopped, Admiral Yi Sun-sin waits, and at Sondolmok on Ganghwa, a strait named after a wrongly accused ferryman, the gatekeeper. The overseas cities follow the same rule: Mizuchi beneath the hidden river under Shibuya, Jenny Greenteeth above London's Fleet River, La Llorona in Mexico City, built on a filled-in lake, and Mae Nak in Bangkok, where canals became roads. Different countries, but buried water and the longing left beside it tie all eight places into one story.
At the center of combat are six guardian spirits: the Azure Dragon, Vermilion Bird, White Tiger, and Black Tortoise, joined by the nine-tailed fox gumiho and the toad. At first only the gumiho is with you, and the others wake one by one as you close gates. Spirit orbs dropped by enemies let you pick blessings such as fire, lightning, frost, water, poison, and charm, and stacking them builds a different fighting style every run. Blessings from two spirits that get along can fuse, like the Vermilion Bird's fire and the White Tiger's lightning combining into a thunderbolt of flame. Each spirit also pairs with one of the six classes: Hero, Warrior, Mage, Archer, Spearman, and Assassin.
Another weakness of the first version was that little carried over after a run. Now you return to the ferry landing. The souls you freed gather on your boat, turning from spirits into stone statues and then gold the more you send on. The souls and bosses you meet become cards, so you can build a deck and play card duels. Many people also wanted to play with friends, so I added two-player co-op.
The third wall: art was too much for one person

I could decide the story and systems on my own, but drawing dozens of characters and maps alone was another matter. PixelLab gave me much better characters than the early ones, yet turning a character changed its look, and animating it twisted its hands or body. Images that looked fine one by one looked like different characters once they moved.

The turning point was Sol arriving in Codex. Handling code and image generation in the same agent environment made me move my work there, and since then I have used Sol, Luna Max, and more recently Astra.

I set a process: define the visual guidelines first, make views of each character from several directions, then produce walking and attacking sprite sheets from those.

To repeat it, I turned the steps into skills and built an Asset Lab and a Map Lab. The mask spirits and guardian spirits new in v2 came out of the same pipeline.

Still, the last awkward pixel or two is something I see best as a designer. So I built a Figma plugin, just for myself, that brings sprite sheets onto the canvas for pixel-level fixes.
Once the v2 key art was drawn, another gap appeared. The illustration looked great, but the actual game screen was not nearly as appealing. Telling AI to "make it prettier" does change something, but without knowing what the difference is, every attempt drifts in a new direction. So I put the key art shown earlier next to real game captures and measured the differences one by one.

There were four main differences. The illustration is framed much closer, and its sunset backlight puts a warm glow on every outline. Orange sky plays against indigo shadows, and layers of sky, distant mountains, lake, and figures give it depth. The game was lit evenly from above, and most rooms sat in a single hue.
Making the whole game look like the illustration would have cost far too much. So I brought the camera close only at the moment that most needs to feel like the illustration, a boss's entrance, and pulled everyday screens toward it with only light and color: a rim light and a sunset grade.


I also decided against some things. When I had Codex repaint the gatekeeper's shading to match the illustration, the shading got richer, but the black outline disappeared and the helmet changed shape. Scaled down to game size, it actually looked muddier. However impressive an AI result looked, if it did not make the game better, it had to go.

I kept a round-by-round record of this work in closing the gap between the illustration and the game.
Maps brought an even stranger moment. I prepared modular assets and asked AI to assemble Seokchon Lake, and it put a shape in the middle of the lake that looked like a raised middle finger. Asked why, it said it had shown only the site of Lotte World for copyright reasons, and that the shape was a bridge leading there. The bridge stopped in the middle of the lake.

Perfecting the modular approach mattered less than releasing the game and finding out whether it was fun. So I reversed the order: generate a full map as a high-resolution image first, then lay collision areas, enemy spawns, doors, and treasure chests over it as vectors.


Changing the art means redoing that layer, but this is still how I make maps today. And Seokchon Lake, the map that gave me so much trouble, ended up as the opening stage of v2.
The fourth wall: the work outgrew my memory
Rewriting the whole story for v2 meant my requests to agents piled up into the thousands. I split the work into sessions for small issues, larger features, and discussing direction with me, but earlier requests still got lost, and sometimes an agent said something was done when it had not been applied.

So I set up a GitHub project and issue board, just as I would with a team. Writing down the situation and spec for each request gave me a record I could return to after the conversation had moved on.
Where I knew less, I leaned on people around me. A developer friend pointed out that the leaderboard, which seemed fine to me, let anyone change scores and nicknames through another route, and I was able to fix it. Building and checking it alone, I would have missed that.
Where The Gate is now

The Gate is now live as v2, with eight gates, six classes and six guardian spirits, the ferry landing and card duels, and two-player co-op. I made one GitHub contribution in 2025 and over 3,000 in the first half of this year alone, and I have been invited to share the experience at several events.

Some problems are still open. Art consistency across directions and movements needs to improve, and someday I want each map asset to carry its own collision data so maps can be assembled as modules.
The biggest lesson of these two years is how clearly the work splits between what AI does well and what I have to decide. Sprite sheets, code, and a plausible world now come quickly. But no one decided for me why the story starts at Seokchon Lake, why it follows buried water, how guardian spirits become combat rules, or what would bring both a struggling player and a skilled one back for another run. Those decisions, piled up, are what The Gate is today.
If you ever walk past Seokchon Lake, think of the Songpa River buried beneath it. And if you play The Gate, try mixing different blessings. I hope you enjoy finding a combination that is yours.

