AI Makes Pokémon

AI Makes Pokémon (experiment / mini-tutorial)

Goal

Build a playable Pokémon-style game in 60 minutes using three high-end AIs and publish it if the result is good.

AI pipeline / workflow

The experiment used a 3-step AI pipeline plus a final check:

1. ChatGPT — Director & base-code generator

   • Referred to in the video as “Chad GBT (5.4 Thinking Model)”.
   • Produced the project plan, basic game architecture, and an initial HTML/JS codebase.
   • Created a simple first-iteration playable clone titled Mini Monster Quest with basic battle mechanics, encounter logic, leveling, and starter monsters (initially no custom images).

2. Google AI Studio with Gemini (3.1 Pro) — Builder & asset generator

   • Took ChatGPT’s code and generated monster images/sprites (examples: Fire Cub, Bubbleoo, Sprattle/Turwig-like).
   • Embedded images into the game code, aiming for a Game Boy / Pokémon Yellow aesthetic and mobile-friendly controls.

3. Claude (Opus 4.6 extended) — Polisher / refactorer / finisher

   • Fixed sprite-loading issues, refactored assets and code, and polished visuals and gameplay toward the intended retro feel.

4. Final check

   • ChatGPT was used again at the end to identify and report remaining bugs.

Technical notes and product features

• Base technology: single-file HTML/JavaScript game (code from ChatGPT was pasteable into one file).
• Asset integration:
  • Gemini generated and embedded images directly in the HTML/JS.
  • Claude produced simpler pixel sprites to better match the retro aesthetic when needed.
• Implemented gameplay features:
  • Wild encounters in grass patches.
  • Turn-based attacks with simple damage calculations.
  • Leveling system and a starter monster.
  • Healing tiles (blue pad), NPC professor/intro text.
  • Mobile-friendly controls and zoom/viewport adjustments to mimic a Game Boy experience.
• Development flow: iterative prompting — generate base code, test, send placeholders to Gemini for assets, then send the updated game to Claude to polish and fix loading/sprite problems.
• Timing caveat: core development steps were initiated within the 60-minute window, but Claude’s extended processing pushed final polishing beyond the strict hour.

Bugs / issues discovered

• Initial build: playable but used text-only/placeholder sprites (no images).
• Gemini iteration: sprites sometimes failed to load, requiring re-running prompts and fixes from Claude.
• Asset originality concerns: reverse image search showed at least one Gemini-generated creature was extremely similar to an existing Pokémon (Turtwig), raising copyright/originality questions.
• Gameplay simplicity and balance: combat mechanics are basic (attack = damage) and balance issues exist (player can die easily).

Outcome

• Final product: a playable Pokémon-style clone titled Mini Monster Quest. The creator published the game and included a link in the video description.
• The creator considered the project successful and plans to continue developing the game further.

Extras mentioned

• Sponsor: a plug for design.com (an AI-powered branding/design tool).
• Personal context: the creator’s fondness for Pokémon Yellow influenced aesthetic and gameplay decisions.

Main speakers / sources

• Video creator / narrator (experiment host)
• ChatGPT (called “Chad GBT 5.4 Thinking Model”) — planner/base code
• Google AI Studio / Gemini (3.1 Pro) — asset & build generation
• Claude (Opus 4.6 extended) — polishing/refactoring