If you tried using AI to speed up your work or get a visualisation of your idea to present, you may have asked an AI to write a document or code a simple prototype.
Often, the result is "fine," but not perfect. It might feel generic, or it might have logic errors. Usually, the problem is the prompt structure.
Each AI model has a different "brain." To get the best results, you need to speak the specific language of the tool you are using.
The Experiment: One Prompt for All
I tested four popular AI tools - Rosebud, Makko, Claude, and Google AI Studio - with the same task: create a 2D top-down puzzle game about a bee finding flowers on a 6x6 grid. I gave them all the same prompt:
"Create a 2D top-down grid-based puzzle game for web/desktop where the player is a bee trying to find specific nectar-filled flowers in a bright, colorful garden. Generate a 6x6 grid of clickable squares that look like patches of dirt. Behind these squares, randomly hide exactly 5 'Target Flowers', exactly 3 'Nets', and fill the rest with a random mix of 'Empty Dirt' and 'Normal Flowers'. Include a UI displaying 'Moves Left: 20', 'Lives: 3', and 'Target Flowers Found: 0'. Implement the following rules: When the player clicks an unrevealed square, permanently reveal the item and subtract 1 from 'Moves Left'. If the item is a 'Target Flower', add 1 to 'Target Flowers Found'. If the item is a 'Net', subtract 1 from 'Lives' and flash the screen red. If the item is 'Empty Dirt' or a 'Normal Flower', just reveal the sprite without changing the score or lives. Finally, add win/loss conditions: If 'Target Flowers Found' reaches 5, pause the game and show a 'You Win!' screen. If 'Lives' reaches 0 OR 'Moves Left' reaches 0 before finding all target flowers, show a 'Game Over: The Bee was caught' screen. Include a 'Retry' button on both screens that fully resets the grid, lives, and moves".
Here is what happened:
- Claude & Google AI Studio: Fast and functional. They gave a working game in one shot, but the design felt like a "template"—very standard and a bit boring.
- Rosebud: It has a 500-character limit for the first prompt, so I had to cut the text down. It works better when you build step-by-step.
- Makko: It struggled with the logic at first and needed several "fixes" to make the game work correctly.
You can see the results here: Claude Design, Google AI Studio, Rosebud, Makko.
The Lesson: One-size-fits-all prompting doesn't work if you want high-quality results.
How to Speak "Model Language"
To improve your output, you must change how you package your instructions for each tool.
I asked Gemini to write a prompt for Google AI Studio, and GPT to write a prompt for Rosebud. This "Meta-Prompting" strategy, asking the parent model to optimize the instruction, significantly improved the game's quality in the second round of testing.
1. Rosebud: The "Step-by-Step" Builder
Rosebud is a specialized game engine. It doesn't like "mega-prompts."
- The Strategy: Use small prompts with one clear action.
- Example: First, ask for the grid. Then, ask to add the bee. Then, add the logic.
- Best for: PMs who want to iterate and "feel" the game as it grows.
- The result of optimised prompt
2. Makko: The "Design Doc" Approach
Makko works best when you organize information like a technical document.
- The Strategy: Separate your prompt into sections: Scene (visuals), Objects (items), Behaviors (rules), and Output.
- Best for: Professionals who want a structured development environment.
- The result of optimised prompt
3. Claude Design: Be Orgenised
Claude is currently one of the best for UI and polished code. It loves structure.
- The Strategy: Use Markdown and clear tags. Give it a Role (e.g., "You are a Senior Game Dev"), a Task, and Constraints.
- Best for: Quick, beautiful prototypes and clean UI.
- The result of optimised prompt
4. Google AI Studio (Gemini): The "Context" King
Gemini processes the top of your prompt as the most important part.
- The Strategy: Put your most important rules at the very beginning. Provide detailed context and give it a few examples of the output you want (few-shot prompting).
- Best for: Complex logic and "all-in-one" generations.
- The result of optimised prompt
Pro-Tip: Use "Meta-Prompting"
If you are struggling to write the perfect prompt for a specific tool, let AI help you.
I asked Gemini to write a prompt for Google AI Studio, and GPT to write a prompt for Rosebud. This "Meta-Prompting" strategy—asking the parent model to optimize the instruction—significantly improved the game's quality in the second round of testing.