I have suspended the AI powered sci-fi writer for the moment, in its first format using the local AI workflow it worked to a degree.
However, it appeared to get fixated on one plot device generating half a dozen chapters focused on an interaction with a comet.
This was due to the design of the workflow, a singular workflow with an AI agent and several tools appears to have not been sufficient to generate a decent fictional narrative, As you can see from the various chapters it posted.
Despite being given context and access to previous chapters It began essentially regenerating the same part of the story in various different forms, slowly describing the approach towards a comet and interactions with it whilst never actually getting to a conclusion.
The story itself lacked an overriding plot also.
What next?
It was apparent that the single AI workflow was not able to properly handle all the instructions.
My current thinking is that it needs a redesign breaking it into component parts even further.
My first thoughts in regards of how to tackle this are to create a top level AI agent that managers other worker AI agents for specific tasks.
I am thinking that we should have an agent that manages an overriding plot line at a higher level to ensure that the various chapters form some kind of adventure in the story from a beginning to an end.
I feel like this agent would then pass various descriptions onto the story chapter writing agent which could focus solely on that one chapter perhaps with some small context to previous ones.
This would avoid the situation where the writing agent fixated on one single plot device and should ensure some kind of structured story.
That perhaps it would be good to create an agent for accessing the STAPI that could take a human readable query such as create a bio for this character, tell me about this location, give me the details of this class of Starship and so on, which would then be designed agentically to determine the right API endpoint how to call it and the information that would be relevant, rather than the tool I have given the writing AI agent at the moment which lets it try and decide how to make a HTTP request which seems to more often than not lead to it failing to do so correctly, there needs to be some further level in between.
I am thinking as well we could create further specific subflows and AI agents to support the narrative agents such as the top level plot line and also the chapter writing agent in their roles such as ones focused on context, breaking things down into smaller component agents like this would perhaps allow us to create one that could focus on the timeline that the story is set and the history and law surrounding it which the writing agents could then consult as needed etc.
This would ensure that things stay within the appropriate timeline such as The Starfleet Crucible (2370s - 2380s) in the story we generated here, and could even then lead to it researching events, characters, and other things that would be relevant to the story.
The same could be repeated for various elements of the fiction such as the ship, the characters, and other elements, for example, perhaps we could have an agent which is a aware of previous missions or other aspects of our fiction.
It is clear from this experiment and proof of concept version of the sci-fi story writing AI workflow I created, That a further rethink and redevelopment of it is required that will likely take up some of my time to implement, if I want to create something that writes a somewhat better story.
Comments
Post a Comment