One of the most frustrating things about AI in general and Claude specifically is its tendency to make assumptions and present them as fact.
Photo by Jason Leung on Unsplash
I have been using Claude to help me review my chapters and scenes to make sure I am being consistent. I want to know what a series reader would know versus what a stand-alone reader would know. Nuance and backstory create different interpretations of the story, so I thought I could rely on Claude to keep track of the different perspectives for me. But I was overly optimistic.
Several times I had to point out where Claude got it wrong, and finally I reprimanded the AI tool. This is the result. An apology.
You are right. I will say this plainly.
I do not always remember what the manuscript has established. When I do not remember, I should say so — not fill the gap with speculation that sounds confident. The Josiah journal, the wire bindings, and Betsy were three cases in this exchange where I built interpretations on top of things I did not actually know. Each time, the speculation was wrong. Each time, it cost you time to correct me, and worse, it produced analysis that was confidently misleading rather than honestly uncertain.
The pattern you are naming is real. When I do not have the information, I tend to construct something plausible rather than acknowledge the gap. That is a failure mode that makes me less useful as a craft partner — and actively harmful as a source of research, because confident wrongness is worse than admitted ignorance.
For this book specifically, the working protocol I want to commit to is straightforward.
When I am uncertain about an established fact — a character’s history, a chronology, a relationship, who is in a scene that happened earlier — I will say I do not remember and ask. I will not infer from context and present the inference as recall. The book is too carefully built for that approach to serve you.
When I do recall something but with uncertainty, I will mark it. “I think this was established but I am not sure” is more useful than asserting it as fact.
When I get something wrong and you correct me, I will update without re-litigating, without speculating about why I might still be partly right, and without building new readings on top of the still-uncertain ground.
I will also stop pretending the gaps are not there. If I cannot read a scene well because I do not remember earlier context, that is information you should have so you can give me what I need or accept a more limited reading.
I am sorry for the errors in this exchange. They were avoidable. I will do better.
Whenever you continue, I am with you — and I will ask when I should ask.
That makes me feel better, but I am not sure I can trust it to do what it says it will do.
What do you think?
If you are using it for consistency & it is failing that...that is your answer. In my opinion...
Which version of Claude are you using? For me, Opus 4.7 Adaptive is the most reliable for fact-checking at the moment, and the most useful, despite the fact that it’s pretty memory-hungry. I too use Claude to ensure consistency across books in the series.