Hi Google AI community,
I have been exploring Google AI Studio and experimenting with building applications using generative AI models.
While prototyping is straightforward, moving an AI application from a demo stage into a production environment introduces several challenges:
- Managing prompts and maintaining consistent outputs
- Handling context and conversation history
- Connecting models with external APIs and data sources
- Evaluating response quality and reliability
- Designing scalable application architecture
I would like to learn from developers who have built production applications using Google AI Studio:
- What practices have helped you improve reliability and response quality?
- How do you manage prompt versions and changes over time?
- What approach do you recommend for integrating Gemini models with existing applications?
Would appreciate any real-world experiences or lessons learned.
You can use the “Copy App” button and give it a different name, or add “Beta” or “Stable” (for stable copies of your apps).
Do some research on using
AGENTS.md
GEMINI.md
PROMPTING.md
CLAUDE.md
.cursorrules
.windsurfrules
.github/copilot-instructions.md
Some of these are pretty specialised and are for specific purposes but can be your friend when used correctly.
Easiest way make an app asking it to explain what these are and how and when to use them. Get it to throw in a few examples as well
Speaking about agents.md…
With my experience with AI Studio, I ask Gemini to implement agents.md file per each folder and subfolder to ensure these directories won’t be removed/modified/pruned/omitted. However; this implementation is not enough. Sometimes; I have to review the folder structure to ensure nothing goes missing as I expand my game/app. I asked Gemini to implement PROTECTIONS_NOTICES.md file at the root of my project to indicate ALL of my crafted code is final and protected. Sometimes… automated processes can break your code by stubbing, deleting, or even simplifying code base… assuming this code is too advanced. I asked Gemini to implement .integrity.json to ensure the integrity of my code is kept in check.
What I just did as I kept adding more capabilities to my game/app; I asked Gemini to implement the Sentinel Integrity System to keep my code base in check, and protect it against automated pruning/deletions/stubbing/zeroing of my code base.
I use the power of computer science as I write prompts that are meaningful.
Sometimes; I have to remind Gemini NOT to edit code just yet, and ask me for verification. Here’s what I used at the end of each of my written prompt:
Don’t edit this game just yet
Description: I ask Gemini not to edit this game just yet.
Pause, ask me for verification.
Description: I use this prompt to ensure Gemini asks me for verification.
Sometimes; I have to use system instructions to ensure AI Studio does its job correctly. If the instructions are meaningful… AI Studio can follow these instructions, if they’re written properly.
Sometimes it is like they have a learning disorder
I am pretty much code illiterate and hundreds of lines of code and snippets in the chat window are gobbydegook to me. So part of my agents said dont show me
When choosing which Gemini model, such as latest; 3.8; 3.7; 3.6; 3.5; 3.1 pro or light; or 3 Flash preview… each model is different, and you should choose a model wisely.
If you are discussing about ideas for game/app that doesn’t require editing code (requires a prompt “Don’t edit this game/app.”)… a lighter model is usually used. However; AI does make mistakes. Always use “Thumbs Down” button to mark mistakes, and clearly write a response to fix assumptions. “Thumbs Up” is intended for correct responses.
If you want to ensure your game/app is doing its job correctly, always test it thoroughly… before you use these buttons what I’ve mentioned earlier.
Always download your code files on a regular basis. This is key to ensure you have proper set of files of your last known files.
Remember: AI Studio by Google should be treated as a tool, not as a burden.
