Earlier today I was thinking about the following seemingly ineffable question, “Assuming that all your knowledge, skills, and abilities were rolled up into a model being served up as a part of an LLM package within an agent, what would your prompts be at the start of the day?” We currently don’t have a lot of unprompted agent interactions. With ChatGPT or other interactions you are generating a prompt. It would be possible to give the model a set of randomly started or scheduled prompt systems. That is a lot like how I start my day where things begin to come into focus and whatever catches my attention ends up being the thing that I write about or consider. That could very well be a flywheel of a prompting system that would allow the agent to interact with you instead of you being the initiator of the prompt.
People may not find it very useful, but I think it is an interesting question about what happens when a model has more of an ongoing conversation or system of prompting to sustain the overall interactive over a protracted period of time. I know that people have created scenarios where agents talk to each other, but what I’m thinking about here is a little bit different. This would be that set of questions or prompts that are the basis of sparking action during the periods of inaction. For me it is the start of the day where I’m just starting to pick a focus or maybe during a period of down time during the day where I go from a few moments of rest to deeply working based on some input.
At the start of the day my basic prompt would be simply, “start writing a page of prose.” That is pretty much how it starts. I open up a Google Doc and start to write about whatever happens to show up in a stream of consciousness until something sticks. Other times I take the other path where I open up my backlog and start working on either the next topic under consideration or one that catches my attention. Those two options are really where the writing starts. To convert that into a prompt it would end up being something like, “Randomly select either writing a page of stream of consciousness based prose in the style of Nels Lindahl or write about the next item in the backlog.” It would be possible to store the backlog as a list for the agent and it could probably actually work that prompt successfully. It would perhaps be better to rewrite that prompt to be, “Randomly select either writing a page of stream of consciousness based prose in the style of Nels Lindahl or write about the next item in the backlog that has not been previously covered.” Expanding that prompt would let you run it each day and not accidentally create coverage of the same backlog items. At this point, without adding that the prompt could very well cause repetitive coverage.
Dr. Nels Lindahl
Broomfield, Colorado