I did something similar [1], where the text based game being a Text Interface (TI) that provides model a view and a set of actions. The model repeatedly interacts with TI to achieve a goal.
In the current implementation, Text Interface allows the model to list files, open a file, and search within a file to find what it needs to satisfy the goal.
With my current set of prompts, the model is able to backtrack when it fails to find relevant info in the current file. However, I couldn't get it work with GPT-3.5. Only GPT-4 is capable to reason its way through the Text Interface.
https://confluo.app - A productivity assistant app that instead of planning your day in advance dynamically suggests you tasks to pick based on the time of the day.
Planning ahead never worked for me as something unexpected always came along. Hence I wanted an app that I can open, go through a list of task suggestions and pick the one I like.
In addition, I included features like timer, pomodoro, and virtual co-working that helped me stay productive during the lockdowns.
I also wanted to track how many hours I was working and which skill I was spending my time on, so I also those features.
I shared it on HN, hoping people will like it. But it never made it to new page :(
That’s how I felt when I was working on my own during the lockdowns. The harder I worked the lonelier I became.
Sorry for the plug, but to resolve this I built confluo.app to keep others informed as we work. I’m still testing this idea out.
In the current implementation, Text Interface allows the model to list files, open a file, and search within a file to find what it needs to satisfy the goal.
With my current set of prompts, the model is able to backtrack when it fails to find relevant info in the current file. However, I couldn't get it work with GPT-3.5. Only GPT-4 is capable to reason its way through the Text Interface.
[1] https://github.com/ash80/backtracking_gpt