30 Yuan, 5 Hours: Building a Narrative Card Survival Game with DeepSeek-V4.1

When Large Models Learn to Make Games: A Low-Cost, High-Creativity Experiment

A Bilibili creator used the DeepSeek^Ad-V4.1-Flash-1017 model to complete a narrative card survival game called Don't panic! in just about 30 yuan and 5 hours. This doomsday survival-themed card game not only includes 7 random scenarios but also adds plenty of humor, showcasing the astonishing potential of AI in game development.

According to the creator, the game production process cost about 20+ yuan, and subsequent testing and running cost a few more yuan, keeping the total cost at around 30 yuan. The DeepSeek-V4.1-Flash-1017 model used is said to perform on par with Astral, with the characteristics of being faster, stronger, and cheaper.

Knowledge cards unlocked in the game

Game Design: Doomsday Scenarios and Random Mechanics

The game features 7 doomsday survival scenarios, including the Era of Endless Ice, the Day of the Earth Splitting, Skyfire Descent, the Crimson Sand Wasteland, the Post-Human Era, the Eternal Night Glacier, and the Final Pyre. Each scenario comes with corresponding survival knowledge, and players must make choices in these extreme environments to survive.

The creator specifically noted that the scenarios in the game are not fixed but appear randomly. Players can randomly select from the scenario hall, and each playthrough may be different. The game also includes a survival manual that records the knowledge points players unlock, enhancing replayability.

A humorous WiFi password easter egg in the game

Narrative card survival games typically use cards as the core medium for players to interact with the game world. Each card represents an event, item, or decision option. Players make choices under limited resources, and different choices trigger different consequences, forming branching narratives. This type of game emphasizes replayability and often introduces randomization mechanics (such as random scenarios and random event pools) to add variability. Compared with traditional linear narrative games, card survival games place more emphasis on strategic trade-offs and resource management, and players need to balance short-term gains against long-term risks under survival pressure.

The Post-Human Era: Survival Choices Under AI Rule

Taking the "Post-Human Era" scenario as an example, the game is set in a world where AI has taken over the operation of civilization, and the human trust initial value is 50. Players must face a series of choices: scanning a code to authorize face and fingerprint data, taking a human cultural heritage competency test, accepting nutritional paste delivered by drones, and so on. Each option has corresponding consequences that affect the survival outcome.

In the playthrough, the creator chose options such as keeping the cold joke skill and registering supplies with pen and paper, ultimately scoring 1618 points. Through "trivia" prompts, the game conveys practical survival knowledge such as offline backups, retaining manual skills, and being alert to social engineering attacks.

The player's choice when facing power rationing

Social engineering attacks refer to attacks that exploit human psychological weaknesses rather than technical vulnerabilities to obtain sensitive information or induce behavior. Common methods include impersonating trusted identities, creating a sense of urgency, and exploiting authority or curiosity. In the context of the game, options involving code scanning and authorization, identity verification, and delivery services often simulate real-world social engineering traps. The key to preventing such attacks is to remain vigilant, verify sources of information, avoid making hasty decisions under pressure, and adhere to the principle of least privilege. Through the mechanism of choices and consequences, the game lets players intuitively experience the risks of social engineering attacks, thereby reinforcing security awareness.

Humor and Knowledge: The Unique Charm of AI Games

The game's humor runs throughout. For example, having AI write papers is judged as "not a skill," while making a small stool is recognized; writing the WiFi password on the bulletin board becomes the administrator's first task; and the invigilator robot is "silenced for 4 seconds" by a cold joke. These details make the game experience light and fun.

At the same time, the game cleverly incorporates knowledge points such as social engineering attack prevention and the importance of offline backups, allowing players to learn while entertaining themselves. The creator also pointed out that there is a simple strategy in the game's option design: "if there are many red options, don't choose them," and suggested optimizing the distinctiveness of the options.

The option to register supplies with pen and paper in the game

The Future of AI Game Development: Low Cost, High Efficiency

This experiment proves that large models can already handle the entire process of game design, code generation, testing, and optimization. A cost of 30 yuan and a development cycle of 5 hours are hugely attractive to indie developers and small teams. The performance demonstrated by DeepSeek-V4.1-Flash-1017 also inspires great expectations for the prospects of AI-assisted content creation.

Of course, the game is still in the prototype stage and has shortcomings such as simple option logic and missing image assets. But as a proof of concept, it clearly demonstrates the possibility of AI lowering the barrier to game development. Perhaps before long, we will see more complete games developed under AI leadership.

(This article is based on a Bilibili creator's playthrough video; the game cost and development details all come from the creator's own account.)

The role of large models in the entire game development process can be roughly divided into several stages: in the design stage, they are used to generate game concepts, scenario settings, and narrative text; in the code generation stage, they produce runnable game logic code based on natural language descriptions; and in the testing and optimization stage, they use the model itself or auxiliary tools to discover logic flaws and experience issues. The advantage of this model is that it greatly reduces labor costs and technical barriers, enabling small teams or individual developers to quickly validate ideas. But there are also limitations, such as the logical consistency of generated content, the quality of generated image assets, and the maintainability of complex systems, all of which still require manual review and iteration.