A few days ago, I wanted to play the mythical DOS game Paratrooper – the one where you control an anti-aircraft gun and shoot down enemy paratroopers. But I didn't want to download an emulator or code it myself. I wanted to see if two different AIs could do it for me, each with a distinct role.
The result was functional, yes, but very basic. No advanced physics, no sound, no helicopters. But the cannon shoots, enemies fall, lives decrease, and you can lose. Just enough to feel nostalgic for five minutes.
How did I do it? Let me walk you through the process, the tools, and the lessons learned.
The two AIs and their roles
| AI | Role | What it did |
|---|---|---|
| DeepSeek (me, the assistant) | Architect / Planner | Designed the project structure (TS, CSS, HTML files), wrote the detailed prompts that would later be sent to the other AI. |
| Nemotron 3 Super Free | Coder | Received DeepSeek's prompts and generated the actual game source code (TypeScript, CSS, HTML). |
Important: I did not use OpenCode with the Nemotron model. I simply copied the prompts that DeepSeek gave me and pasted them into the Nemotron 3 Super Free interface (a free model that claims to understand legacy code and retro languages, available on several AI platforms).
Step 1: DeepSeek plans the structure
I asked DeepSeek to help me create a Paratrooper game in TypeScript, with CSS and HTML, but with one condition: the code had to be generated by another AI. So DeepSeek didn't write the game – instead, it generated a series of ready-to-copy-paste prompts.
These prompts included:
- A
tsconfig.jsonto compile TypeScript. - An
index.htmlwith canvas, life counter, and restart button. - A
styles.csswith a retro look. - A full
game.tswith the game logic: mouse-controlled cannon, click-to-shoot, falling paratroopers, collisions, lives, game over, and restart.
DeepSeek also warned me that the result would probably be very basic if I didn't iterate much. It was right.
Step 2: Nemotron 3 Super Free writes the code
I copied the longest prompt (the one for game.ts) and pasted it into Nemotron 3 Super Free. In less than a minute, the AI returned complete code. Then I did the same for the other prompts.
When I opened index.html in my browser… it worked on the first try! The cannon moved with the mouse, paratroopers fell from the top, shots eliminated them, and lives decreased when they hit the ground.
But I also noticed clear limitations:
- Imprecise collisions – Sometimes a bullet passed near an enemy and still killed it.
- No fire rate limit – I could shoot like a machine gun.
- No animations – Paratroopers were just brown rectangles with a white square on top (the parachute). No sprites.
- No helicopters – Just basic paratroopers.
- Restart sometimes left "ghost bullets" in the array.
In short: it delivers the core mechanics, but without any refinement. It's a functional prototype, not a polished game.
Can you improve it with more prompts?
Absolutely. I asked DeepSeek to generate correction prompts: limit shots, adjust collisions, clean the array on restart. I copied those new prompts to Nemotron, and after two or three iterations the game improved noticeably.
But even after several rounds, the result was still "very basic" compared to the original DOS Paratrooper. The reason? Nemotron 3 Super Free seems optimized for simple, straightforward code, not complex simulations or games with delicate state.
What I promised vs. what I got
| Feature | Final state |
|---|---|
| Cannon follows mouse | ✅ Yes |
| Click to shoot | ✅ Yes |
| Paratroopers fall | ✅ Yes |
| Bullet-enemy collisions | ✅ Yes, but somewhat imprecise |
| Lives (3) and game over | ✅ Yes |
| Restart button | ✅ Yes, though sometimes buggy |
| Fire rate limit | ❌ Not initially, yes after correction |
| Sound | ❌ No |
| Explosions | ❌ No |
| Helicopters | ❌ No |
| High scores | ❌ No |
In one word: basic.
Lessons learned (if you want to try this yourself)
- AIs don't replace human judgment – Without my corrections (the extra prompts from DeepSeek), the game would have been almost unplayable.
- Nemotron 3 Super Free is great for rapid prototypes, but don't expect a commercial game.
- DeepSeek (or any conversational assistant) is ideal for designing architecture and prompts, because it understands the overall project context.
- Iteration is key – With each correction prompt, the game gets a little better. Three or four iterations already give a decent result.
- "Very basic" isn't bad – Sometimes we just want to revive a memory, not create the next indie hit. And for that, these tools are perfect.
Final reflection: was it worth it?
Absolutely. In less than an hour, without writing a single line of manual code, I had a playable game in the browser. My 10-year-old nephew tried it and had fun shooting at the paratroopers. Sure, after five minutes he said "it's too easy" and went back to Roblox. But for me, seeing that cannon move clumsily on screen was like traveling back to 1992.
If you also want to revive a DOS classic without the hassle, I recommend this method: use one AI to plan (DeepSeek, ChatGPT, Claude) and another to code (Nemotron, Copilot, Codeium). The result will be basic, but functional. And the fun is in the process.
Bonus: the (very basic) final code generated by Nemotron
In case you want to try it or improve it, here's the link to my repository:
➡️ https://github.com/davdomin/paratrooper
It's not pretty, but it works.
Have you tried building games with two different AIs? Share your experience in the comments! 🕹️
Top comments (0)