My latest deep dive into using frontier AI models to research and write scholarly papers — this time with the just-released Fable 5.1 model from Anthropic
I absolutely love this. 2 questions: 1) Can you share your setup with Claude code so we can try to replicate your experiment in an area where we are knowledgeable? 2) Have you tried to take a paper written by humans that was rejected and use AI to make it meet the requirements for publication? I think that would be a more interesting project than an AI authorship (although I love what you have done).
1. It was a conversational process — I always find these more productive than a rigid workflow — but starts with setting up a Claude Code project in a folder, talking with Claude about what I want to achieve, letting it know that I will retrieve digital copies of papers it can't access directly, and asking it to set to a folder structure, project workflow, and plan, for this particular case. At some point within the workflow I will provide detailed comments on drafts (either in-line comments in markdown files or comments added to a Word file), and request adversarial reviews. In the past I have spent some time working within the project to establish an author writer profile, but this has had limited success.
The clever thing though is that Claude Code is able to work with you on the best project setup and workflow for the task.
2. I haven't yet but I know others have — and this seems to be a no brainer, as Claude Code works much better with starting material and clear objectives. And alwasy in Claude Code (or Codex) and not a browser session!
I've heard that Fable is not an especially good writer. There is part of me that wonders if this is a byproduct of increasing these models ability to code or (more sneakily) a deliberate attempt to actually downgrade their models abilities to write to circumvent all the consternation around AI writing. Opus is still very impressive (and my default) but I have not seen major improvements in its writing over the past 6 months not withstanding the serious increase in capabilities on the agent and coding front.
I absolutely love this. 2 questions: 1) Can you share your setup with Claude code so we can try to replicate your experiment in an area where we are knowledgeable? 2) Have you tried to take a paper written by humans that was rejected and use AI to make it meet the requirements for publication? I think that would be a more interesting project than an AI authorship (although I love what you have done).
Thanks Beny.
1. It was a conversational process — I always find these more productive than a rigid workflow — but starts with setting up a Claude Code project in a folder, talking with Claude about what I want to achieve, letting it know that I will retrieve digital copies of papers it can't access directly, and asking it to set to a folder structure, project workflow, and plan, for this particular case. At some point within the workflow I will provide detailed comments on drafts (either in-line comments in markdown files or comments added to a Word file), and request adversarial reviews. In the past I have spent some time working within the project to establish an author writer profile, but this has had limited success.
The clever thing though is that Claude Code is able to work with you on the best project setup and workflow for the task.
2. I haven't yet but I know others have — and this seems to be a no brainer, as Claude Code works much better with starting material and clear objectives. And alwasy in Claude Code (or Codex) and not a browser session!
I've heard that Fable is not an especially good writer. There is part of me that wonders if this is a byproduct of increasing these models ability to code or (more sneakily) a deliberate attempt to actually downgrade their models abilities to write to circumvent all the consternation around AI writing. Opus is still very impressive (and my default) but I have not seen major improvements in its writing over the past 6 months not withstanding the serious increase in capabilities on the agent and coding front.
I wonder that as well
Journals are already grappling with this — and this is where it's worth reading Robert Braun's article