The general idea of this project is to automate tasks using a Cohere LLM to create applets, save these applets and let them be used in a user-friendly way.

Step 1: Automate tasks with using code generation

  • Examples:
    • “Text Zara and ask what I missed in class today and email TA about absence“
    • “Make a GUI to with a counter that can be incremented/decremented”
  • Prompt is fed to LLM, which writes code and test cases for itself
  • Iterative prompting for self-debugging: if there are errors or failed tests, the LLM prompts itself again with the original prompt AND the error message
  • Once errors are resolved, the applet prototype is shown to the user for approval
    • User can give feedback like “this is not what I want it to do” or “please make the GUI pink instead”, at which point the LLM will prompt itself again with (original prompt + code + feedback) to keep improving the applet
    • If the user is happy and has no feedback, the applet is finalized
  • Small features that are easy:
    • Automatically name applet from prompt
  • Nice-to-have fancy features:
    • Virtual environment creation with dependencies for each applet’s needs
    • Fancy skill-building shit (Voyager)

Step 2: Applet Dashboard

  • Since the applets are just code generated by LLM, they are easy to save – just save the code into a directory
  • Make sure code is in executable format (easy with Python code)
  • For every code file/applet in the directory, make a button on a dashboard; when the user presses the button, the applet runs
  • Dashboard can be a web app or a desktop app, maybe even mobile

Step 3: Sharing

  • Crowdsourced app store platform (social media???)
  • If you create an applet that you like and think other people would find useful, you can upload it to the platform
  • Then, other users can download applets from the platform
  • (This is just GitHub basically lol)
  • Nice-to-have fancy features:
    • Semantic search of applets in the store

Some notes:

  • The bottleneck of what our system can do is LLM capability
    • so if stuff doesn’t work we blame Cohere :)
  • Some bullshit statements to sprinkle in:
    • “Because your generated code is saved locally, it’s personal and secure”

Key Features

  • Generate an applet from user prompts
    • Code + test cases
    • Self-debugging and testing
    • Name and icon generation
    • Reverse prompting
  • Save your applets to be re-used
  • Share your applets with other people