I’ve been experimenting a bit with this and I’m just feeling that this system could potentially be a lot better with something like a multi-agent setup.
If I want to write an epic sci-fi novel, it would be very nice to give the characters their own contexts and similar. I was thinking of using something like langgraph or similar just to set this up but I’m not sure if it’s needed?
Do you find the characters blending together or general styles of thinking blending together? How good is it at world building and holding facts about the world consistent over time?
I experimented a little bit with a multi-agent setup, though instead of the agents being different characters, I had two different writer personas. One was prompted to be more focused on the details of the scenes, while the other paid more attention to the overall story structure and to ensuring that there were enough open loops. I put together a simple Python script that had them talking with each other and which allowed me to edit each message after it was generated, to fix any obvious issues. Adding something like a dedicated critic to automatically check each message for logical consistency etc. would also make sense, though I didn’t try it as I just acted as the critic myself.
I thought the results were interesting, but it required using the API rather than the web interface, so the token costs went up fast and I haven’t pursued it beyond a few experiments. It also had a different feel to it—more like I was a director rather than a co-writer. Which is not necessarily bad! If it was cheaper, I’d probably experiment with it more.
If you do go for a multi-character setup, something like Silly Tavern might be a useful tool (I haven’t looked into it much myself).
I feel like Claude is pretty good at worldbuilding as long as you think to ask it about something, but it’s usually limited in what non-obvious implications it comes up with unprompted. With my style of writing, it seems to keep track of details quite well, though since I also write a big chunk of the text, I keep actively reinforcing things like setting details and character voices. I expect that it would do much worse if it was running purely autonomously.
I’ve been experimenting a bit with this and I’m just feeling that this system could potentially be a lot better with something like a multi-agent setup.
If I want to write an epic sci-fi novel, it would be very nice to give the characters their own contexts and similar. I was thinking of using something like langgraph or similar just to set this up but I’m not sure if it’s needed?
Do you find the characters blending together or general styles of thinking blending together? How good is it at world building and holding facts about the world consistent over time?
I experimented a little bit with a multi-agent setup, though instead of the agents being different characters, I had two different writer personas. One was prompted to be more focused on the details of the scenes, while the other paid more attention to the overall story structure and to ensuring that there were enough open loops. I put together a simple Python script that had them talking with each other and which allowed me to edit each message after it was generated, to fix any obvious issues. Adding something like a dedicated critic to automatically check each message for logical consistency etc. would also make sense, though I didn’t try it as I just acted as the critic myself.
I thought the results were interesting, but it required using the API rather than the web interface, so the token costs went up fast and I haven’t pursued it beyond a few experiments. It also had a different feel to it—more like I was a director rather than a co-writer. Which is not necessarily bad! If it was cheaper, I’d probably experiment with it more.
If you do go for a multi-character setup, something like Silly Tavern might be a useful tool (I haven’t looked into it much myself).
I feel like Claude is pretty good at worldbuilding as long as you think to ask it about something, but it’s usually limited in what non-obvious implications it comes up with unprompted. With my style of writing, it seems to keep track of details quite well, though since I also write a big chunk of the text, I keep actively reinforcing things like setting details and character voices. I expect that it would do much worse if it was running purely autonomously.