Prompt Wars: Human vs AI

• 

  Neon-themed “Prompt Wars” interface showing a duel between two prompts in a futuristic UI.

• 

  Game screen with glowing neon buttons offering different prompt battle modes and difficulty options.

• 

  User entering the prompt “an astronaut riding a horse on Mars” in a neon-styled input box.

• 

  System processing prompts with animated text indicating generation and comparison in progress.

• 

  Side-by-side AI-generated images comparing two prompts, showing an astronaut riding a horse on Mars.

• 

  Victory screen showing the winning prompt result in a neon-styled interface.

Out-prompt the bot. Best image wins.

A neon-arcade, mobile-first game where a human and an AI bot face off in 3 rounds of prompt-crafting. Each round, both players get the same secret theme, write a prompt, and an image model generates their entry. The better image wins the round. Best of 3 takes the match.

💡 Inspiration Everyone's learning to "prompt" — but nobody really knows if they're any good at it. We wanted to turn prompt engineering into a sport: fast, visual, competitive, and playable on a phone in 60 seconds. Think Street Fighter, except your weapon is a sentence and your fists are diffusion models.

The arcade neon vibe came from the idea that prompting feels like casting spells into a black box — so the UI should feel like a CRT cabinet humming in a dark arcade.

🛠 How we built it Framework: TanStack Start (React 19 + Vite 7) for SSR + server functions Backend: Lovable Cloud (Postgres + Auth + Edge Functions) for match state and round persistence AI: google/gemini-3-pro-image-preview via Lovable AI Gateway → generates both the human's and bot's images google/gemini-2.5-flash → judges the two images against the theme and returns a winner + reasoning openai/gpt-5-mini → powers the AI opponent's prompt-writing brain (with difficulty tiers: Rookie / Pro / Grandmaster) Design: Custom OKLCH neon tokens (--primary hot pink, --secondary cyan) with text-glow-* utilities and scanline overlays in src/styles.css Mobile: Tap-only controls, large hit targets, no downloads, runs in any mobile browse What we learned: Latency is the game. Image gen takes 6–12s — so we made the "waiting" screen part of the fun (CRT static, taunting bot dialogue). Judging models hallucinate winners if you don't pin them down. We fixed this by forcing structured JSON output with explicit scoring rubrics (composition, theme-fit, creativity). Difficulty = prompt quality, not model swaps. The bot uses the same model at every difficulty; what changes is its system prompt — Rookie writes 4 words, Grandmaster writes cinematic 60-word descriptions. Mobile-first kills clever desktop layouts. We rebuilt the arena 3 times before it felt right at 375px wide. Challenges: Race conditions between the human's submit and the bot's auto-submit — solved with a server-side round lock. Cost runaway during testing (image gen is expensive). Added per-session rate limits. Making the AI feel like an opponent, not an oracle. Gave it a personality (BIT-9), a win/loss counter, and trash talk that escalates with difficulty. Fitting a 4-step game loop into one thumb-scroll without it feeling cramped. AI tools used: Tool Purpose Lovable Built the entire app (frontend, server functions, DB schema, deploy) Lovable AI Gateway Routed all model calls — no API keys needed Gemini 3 Pro Image Round image generation Gemini 2.5 Flash Judge model GPT-5 Mini Bot opponent's prompt writer Lovable Imagegen Hero/thumbnail art

Built With

• lovable