My first game, developed and released in just 4 days using Gemini and Codex (within Cursor), is detailed below: It's a swipe-based dating simulation game that blends survival management and narrative choices. While the game was developed in just four days using Gemini and @cursor_ai, I found that development speed depends entirely on strict architectural guidelines, not just on AI. 1/ Project structure is crucial. To prevent the AI from cramming all logic into a single, unmanageable file, establishing a clear documentation structure from the outset is essential. I used Gemini to divide the project into four distinct parts: User Interface (UI), Game Logic, Content/Text, and Display Logic (HTML). This separation made the code easier to migrate to Cursor and allowed for seamless modifications and additions later. 2/ Let AI write the code, not the story. Directly involving AI in narrative creation often leads to a significant decrease in efficiency, as the model takes longer to process large amounts of code mixed with story text. Furthermore, AI often struggles with narrative coherence and pacing, frequently producing overly dramatic or abrupt transitions that clash with human storytelling styles. By focusing AI on logic and manually filling in the narrative content, I gained better control over the game's performance and pacing. 3/ Be Cautious with Data While AI is an excellent tool for validating statistics, it shouldn't be relied upon to independently generate core game data. AI often fails to understand the actual impact of events in the game—for example, it might deduct lives because it identifies positive social actions as a general "cost." However, AI is invaluable for numerical auditing; I use it to examine the overall distribution of rewards such as "charisma," calculating medians and averages. This allows me to quickly pinpoint and fix balance issues. 4/ Development Order is Crucial: Numerical Analysis → Narrative → Art The rapid feedback from AI can easily overwhelm players with fun but unnecessary features. I prioritized the attribute system because even without a story or visuals, the core functionality of the game must be robust and engaging. After validating the numerical loop with a simple Minimum Viable Product (MVP), I expanded the project from an initial 10 general events to nearly 100 events across three phases, ensuring its foundation remained solid as the system became more complex. 5/ System Design You Can Trust AI For. While I prefer starting with minimalist systems, I frequently consult AI to resolve bottlenecks during game expansion. For example, when an NPC's refusal might disrupt narrative flow, I use simple prompts to describe the player's expected experience. These suggestions are insightful—for instance, Gemini suggested a "save system" for players with low attribute values.