I've been working on a project called 'Magi', which aims to create an AI system that improves its own prompts. The core idea is to use a recursive meta-process where the AI analyzes its output and upgrades the prompts accordingly.

Key features:

- Recursive meta-process

- Multi-character system for diverse analysis

- Attempts at self-evaluation and optimization

Here's a sample of how it works:

Initial prompt: "I want to be rich"

Final prompt (after 5 iterations): "I'm looking for data-driven, innovative strategies to become wealthy. Can you provide optimized approaches in terms of financial planning, investment methods, and self-development that can be applied immediately? It would be greatly helpful if you could explain the data analysis for each method, describe their long-term impacts, and provide intuitive insights through successful case studies."

While the system does show some improvement, I'm facing several challenges:

1. Character personas tend to manifest in superficial expressions rather than generating new ideas

2. Achieving meaningful and relevant prompt improvements

3. Establishing objective self-evaluation criteria

4. Preventing the system from going off-track or becoming too verbose

The ultimate goal is to evolve this into a universal framework capable of improving all kinds of input-output processes (for example, this could even include improving Magi's own codebase).

However, I feel like I'm hitting a wall with the current approach.

I'd greatly appreciate your insights and suggestions:

- How can I improve the system's ability to generate meaningful upgrades?

- Are there alternative approaches I should consider?

- Do you have any recommendations for effective persona prompts that could lead to more diverse and innovative ideas?

TThis process is open-source on GitHub: https://github.com/ParallelKim/Magi

Thank you in advance for your thoughts and feedback!