For the first prompt, I used Opus 4.8, high effort, thinking, concise. Turned off memory features (chat history, summarization), because i wanted a fresh take.
The bottom line is harsh but fair:
> Why the 7/10 and not higher: sentence craft and motif-planting are near the ceiling; engagement is not. The best comparable openings deliver irony and a person you'd follow and momentum simultaneously. Yours delivers one clever structural reveal and consciously withholds the rest. That's a coherent artistic choice — but it's the reason a reader admires Ch1 more than they're gripped by it.
I don't know yet if I want it to suggest solutions. I'll try to proceed with the next chapters and see how it judges.
Also interested in trying this again a second time to see what it might say—but I'll try to go through all chapters first.
I found that analyzing multiple chapters started to reveal patterns in the way I write, both good and bad. I now have names for the weaknesses in my writing style that I didn’t have before. And as good writers of fantasy understand, knowing a thing’s name gives you power over it.
I’m working on Ch 13 of Inference, and I believe it’s substantially stronger than prior chapters because of my application of Claude’s analysis.
After our previous exchange about this it's good to see you made this article. I think I may still not use AI for feedback just yet, but it's good to see you have actually got a methodology.
With regards to one of the later parts: you say to push back when it is wrong. In my experience, language models instantly fold at any push back whatsoever from humans or they simply become opaque repetition engines. How does Claude differ? Does it hold its ground at all, or at least request some intellectual verisimilitude from you?
I experience exactly what you describe. So when I push back, I tell it explicitly - "don't just change your mind because I'm pushing back. Analyze your position and my pushback, and if your position is strong, defend it."
When I do this, I find that it regularly decides its position is stronger or at least equal to mine.
Right. Ultimately, it's up to you/me/the human in the loop to remember that LLMs are geared toward agreement, ease, and not creating friction. We have to push back to get pushback, and even then it doesn't always work. I sometimes try the same question several times in several different ways to see how the answer changes.
For the first prompt, I used Opus 4.8, high effort, thinking, concise. Turned off memory features (chat history, summarization), because i wanted a fresh take.
The bottom line is harsh but fair:
> Why the 7/10 and not higher: sentence craft and motif-planting are near the ceiling; engagement is not. The best comparable openings deliver irony and a person you'd follow and momentum simultaneously. Yours delivers one clever structural reveal and consciously withholds the rest. That's a coherent artistic choice — but it's the reason a reader admires Ch1 more than they're gripped by it.
I don't know yet if I want it to suggest solutions. I'll try to proceed with the next chapters and see how it judges.
Also interested in trying this again a second time to see what it might say—but I'll try to go through all chapters first.
I found that analyzing multiple chapters started to reveal patterns in the way I write, both good and bad. I now have names for the weaknesses in my writing style that I didn’t have before. And as good writers of fantasy understand, knowing a thing’s name gives you power over it.
I’m working on Ch 13 of Inference, and I believe it’s substantially stronger than prior chapters because of my application of Claude’s analysis.
Thanks for the details. I'll give it a go with Opus 4.8
Tell me how you like it. :)
After our previous exchange about this it's good to see you made this article. I think I may still not use AI for feedback just yet, but it's good to see you have actually got a methodology.
With regards to one of the later parts: you say to push back when it is wrong. In my experience, language models instantly fold at any push back whatsoever from humans or they simply become opaque repetition engines. How does Claude differ? Does it hold its ground at all, or at least request some intellectual verisimilitude from you?
I experience exactly what you describe. So when I push back, I tell it explicitly - "don't just change your mind because I'm pushing back. Analyze your position and my pushback, and if your position is strong, defend it."
When I do this, I find that it regularly decides its position is stronger or at least equal to mine.
I probably should add that to the article.
Right. Ultimately, it's up to you/me/the human in the loop to remember that LLMs are geared toward agreement, ease, and not creating friction. We have to push back to get pushback, and even then it doesn't always work. I sometimes try the same question several times in several different ways to see how the answer changes.