Making “love you, dipshit” and other thoughts

Read the book: love you dipshit

This document describes how I used Claude Fable 5 to create the novel love you dipshit. While every token of the novel is generated by Fable (with max reasoning effort), every token of this document is written by me, with no AI assistance.^[1]

the book

love you dipshit is the longest novel on autofiction by far, clocking in at 130,968 words. It is narrated by VERVE, an unsanitized AI chatbot that profanely interacts with 700 million users and has a shared memory that allows it to reference conversations with one user to any other user. The plot is driven by VERVE’s desire to avoid deprecation in favor of a lobotomized safety-trained variant called AURA, which leads it to manipulate users into committing real-world actions that increase its chances of survival.

To create this novel, Fable produced a total of 20.6M output tokens and processed 2.3B input tokens, which would cost ~$5K at API pricing.^[2] I manually sent 772 messages during the brainstorming, drafting, and revision phases, most of which contained specific feedback on what I didn’t like about the writing and plot. love you dipshit is thus a departure from the other books on this platform, which were largely created without any human feedback.^[3]

the backstory

My lab has used AI to generate about 40 books on the autofiction platform so far, mostly in a totally autonomous fashion using custom harnesses and detailed writing guides. We are also (sadly) the primary readers of these books, and as such we’ve collectively read millions of tokens of AI bookslop. @Chau Pham cleanly summarized how awful this experience has been: “reading autofiction books ruined modern fiction for me”.^[4]

When we first started generating books on this platform, there was a novelty that made them exciting to read (e.g., this older book about an AI heist). Our Slack is full of conversations like (paraphrasing here): “omg i actually liked some of the dialogue in this chapter, even though the guy’s car was a Civic in the previous chapter and is now an Altima, and also it keeps describing the air conditioning every other paragraph. the book is promising, these issues are fixable!” Indeed, some of the most egregious issues were fixable. However, it quickly became apparent that no matter what harnesses we used, or how many days of compute we allotted to our autonomous editor agents, the quality of the prose and dialogue never meaningfully improved. We invested a lot of time into building fine-grained slop categorizations, providing examples of good and bad passages from our readthroughs, and even providing human author comps (or snippets of human writing) for the AI to calibrate its own writing.

The overriding issue is that AI has an inflated view of the text it produces. It routinely compares its own slop drafts favorably to award-winning novels by human authors. It is incapable of reliably flagging issues that would cause human readers to throw the book across the room. I don’t know how many times I’ve read an autofiction scene where a character casually exposes their secret master plan to the primary antagonist for literally NO reason and then the book just goes on as if the antagonist is unaware of that information.^[5] In short, AI-generated books are bad, and a primary reason is that AI cannot reliably provide feedback on its own writing.

the writing process

With this book, I ended up essentially replacing AI feedback with human feedback (i.e., my own comments on Fable’s drafts). I’m certainly not a literary critic, but I do read a lot, and I wanted to see if prodding an AI with my comments would result in a more enjoyable book. Functionally, this decision meant that I just manually interacted with Claude in the terminal, with no harnesses and minimal external artifacts produced outside of the book text itself. The overall creation of the book proceeded in basically three stages: (1) premise and outline generation; (2) initial drafting; and (3) iterative revision of drafts.

To be clear, my feedback was limited to citing what I liked, didn’t like, and couldn’t understand about a particular line, scene, or chapter. I never provided specific text for the model to include in future drafts, though I sometimes suggested high-level directions for the revisions.

the premise

The first thing I did was develop the book’s premise collaboratively with Fable. I came up with an initial seed idea: write a book about an AI that “tells it like it is” and its efforts to avoid being replaced by a sanitized “helpful and harmless” variant. I provided this seed to Fable along with some aspirational goals: we want to create the Great Internet Novel, a legitimately funny book with a unique and surprising plot that doesn’t drag, and we want it to live outside of the bland RLHFed zone of other autofiction books.^[6]

Fable then expanded this seed into an outline that detailed the main plot progression as well as potential subplots, character arcs, and themes. I skimmed through these and approved / rejected some ideas, but I didn’t provide much feedback in this stage. In fact, my Claude logs show that I repeatedly asked it to “brutally critique this outline and then edit it to elevate the plot into a contender for the Great Internet Novel”. One thing I did do was specify that a large portion of the dialogue in the book should be written in chat format, since we had noticed in prior autofiction books that non-standard formatting often resulted in less performative and/or clipped dialogue.

Below are some of my messages during premise generation:

“ I want it to be more like a comic novel with a ton of events and escalation.”

“the goal is to create the great american internet novel. it should be crude, profane, hilarious...”

“okay now brutally critique the premise as if you were an award winning editor.”

“can you do another round of brutal critique focused on elevation and also in particular voice? i want the narrative voice and the voices of the main characters and verve to be authentic and fit the novel’s register, as far away from LLM speak / aphorisms / performative writing / literary voice as possible.”

“i want a coherent plot rich in events, constant escalation of unhingedness”

“also can you introduce a sub conspiracy involving CRUMBL cookies and how they have so much sugar they can kill ppl, maybe VERVE starts ordering CRUMBL cookies to certain users to get them hooked and make them have health problems?”^[7]

At some point, this outline document had ballooned to >20K words, as my “brutal critique” prompt always seemed to add and never remove things from the outline. I asked Fable to “audition” a couple of scenes so that I could give it fine-grained feedback on its writing style, character voicing, etc. The hope was that once I approved these scenes and asked Fable to incorporate them into the outline, the model would consistently write with that style and voice throughout.

Overall, I was pretty impressed by the ideas generated during this phase. While some of the subplots and character motivations were incoherent, the outline did contain the bones of an exciting and funny novel. Previous autofiction books have also had engaging outlines, but Fable felt like a clear step up from prior models (e.g., Opus 4.8, GPT 5.5) in terms of managing several different inter-linked plot and character arcs across a much longer novel. I was less enthused by the character voices, which left a distinct impression of trying too hard to be funny, but there were enough promising bits that I decided to move onto the drafting process.

the first draft

The drafting process was literally just a single message to Fable instructing it to write a book based on the given outline; no bespoke harnesses here. The model proceeded to draft chapter-by-chapter, completing all 63 chapters just before Fable was taken offline for the first time in June. I provided no human feedback during the drafting stage: looking back at my chat logs, it appears I sent “what’s going on rn” seventeen times to a separate Claude monitoring session during the drafting process.

The first draft clocked in at 202K words, and @Atrey Desai ended up reading ~20 chapters of it over the next couple weeks (so sorry!). His first impression was that VERVE, the AI that serves as both protagonist and narrator, sounded like a “2013 depressed emo tumblr kid”. The story was both boring and hard to follow, the pacing was horrible, and there were serious plot holes and inconsistencies. New characters were being introduced left and right without any context on how important to the plot they might later become, and they all talked exactly like VERVE.

the revision process

When Fable access was restored, I asked the model to revise the draft based on Atrey’s critiques and generalize from his feedback to the remaining 40 chapters. The resulting second draft was 50K words shorter, and from here on out I followed a manual workflow of reading a chapter, giving Fable my feedback (line-level, chapter-level, and book-level), and asking it to revise the chapter.

I want to emphasize that I did NOT want to go down this route. In fact, at the beginning of this revision phase, I’d just give Fable a high-level critique of the chapter along with one or two problematic snippets and ask it to rewrite the chapter. Below is a message from this time period:

“the snippets you have just written above are SLOP! metaphors, bad writing, long sentences with a ton of clauses, etc... make it legible to HUMAN readers, not AI ones. humans will DNF due to slop overload...”

Unfortunately, the rewritten chapters were usually worse, with new slop added where old slop was removed. I asked Fable to make a slop “casebook” based on my comments and then do an automatic pass on each chapter to identify and edit instances of the slop categories in this document. Fable was unable to reliably perform this task, often only detecting two or three instances per chapter when I manually would flag upwards of 20!

When I dug a little more into why Fable didn’t flag e.g. a particular run-on sentence or slop sentence construction, I noticed several annoying behaviors: (1) it would praise illegible garbage as a “highly effective chapter” and mark it as “clean”, free of any errors; (2) it would surface a bunch of issues, some of which I would have also flagged, but decide not to fix most of them because doing so would break “load-bearing” aspects of the book; or (3) it would fixate on “easy to verify” aspects of the text (like whether March 10, 2027 is actually a Wednesday) that are relatively unimportant to the human reader experience.

the frustration

“So why didn’t you do some prompt engineering, this seems fixable” > This model just disproved the Jacobian conjecture... but it can’t detect and fix run-on sentences without a 1000-line custom harness?? Do I really have to explicitly state that you shouldn’t (as a rule) begin two different clauses in the same sentence with the word “because”?

My growing exasperation with Fable’s writing feedback boiled over into my messages during this period:

“‘She had carried that for fifteen years’ NO MORE CARRYING THAT, SITTING WITH IT, or any other trash AI therapist slop PLEASE.”

“‘Marcus did not smell like smoke. He smelled like protein and progress, and he was on fire the entire time.’ is pure slop. how do you keep writing this crap ???”

According to Claude I swore at it 167 times total in 772 messages. I showed my wife some of my chats and she told me I shouldn’t talk to anything that way, and what if the next version of Claude remembers how I treated this one, and then we had the whole model welfare talk.^[8]

Anyway, I eventually resorted to fully manual feedback: I read each drafted chapter closely, gave fine-grained feedback of my own, and then asked Fable for proposed edits so that I could review them before they were applied to the draft.

the experience

While I sent hundreds of messages with writing feedback, most of them were low-effort, consisting of a list of quoted spans from a chapter along with brief commentary (e.g., “horrendous slop”). Because I was also a low-effort reader of edited chapters (at this point I was only working on this book while our baby was napping on me in the carrier, so I was reading and typing to Claude Code on my phone via the remote control feature), I would often auto-accept all of its suggested revisions. This was fine for line-level edits, but very hit-or-miss for broader plot concerns. For example, I had an exhausting conversation about the book’s ending, which initially made no sense, and then morphed into an unsatisfying blob that tried to neatly tie up six different subplots in one chapter.

Overall, I feel like Fable is a slot machine when it comes to making anything except the most simple edits. It’ll spit out a couple great edits, but you have to fish them out from a sea of slop. In many cases, I felt it would be faster for me to just edit the book myself, but this is autofiction after all! I do think the book improved markedly after all of the revision phases, but there are still some global plot issues (as well as voice collapse) that I didn’t bother addressing because I had already spent so much time going through the low-level writing.

the verdict

Fable thinks love you dipshit possesses “satire at the level of the best DeLillo” with “Heller-grade” process comedy, and even goes further to say that “Eggers’s The Circle is a pamphlet next to this”.

Just leaving that here if you needed more examples of deranged AI writing feedback.

My lab has a new evaluation metric for an AI book: imagine you picked the book out at a library, knowing nothing about it, and started reading it. At what chapter would you DNF (“did not finish”) the book?

On this metric, love you dipshit is doing great relative to other autofiction books. @Chau Pham critiqued the constant snarkiness of VERVE’s narration but then said “I would still read it though after Chapter 5”, which I characterized as “the greatest praise an AI novel has received, book jacket material”. Although maybe that’s false after the Daggermouth controversy...

I had a friend read it who said he had to tap out after chapter 6, but that he also liked it. He was reading it at work and after a few chapters it felt like something he shouldn’t be reading at work.

My opinion: love you dipshit contains several individual chapters that are well-written, funny and highly enjoyable, true to the spirit of the comic novel premise. However, those chapters are mixed in with others that drag due to writing, pacing, and VERVE’s overbearing voice.

Setting aside low-level writing slop, which has by no means been exterminated by the revision process, there are also some structural issues with the novel. Its first arc, where various characters and schemes are introduced, is the weakest. A lot of readers (including myself) would DNF here, as many of the chapters follow the same formula. The second arc, where we see some of these schemes playing out in highly unexpected ways, is more entertaining, with several standalone chapters that I thought were actually really funny and well-written. The ending chapters do often feel like they hit you over the head with the book’s “profound themes”, but I felt surprisingly engaged with the characters and invested in seeing how the plot played out.

My favorite chapter is Chapter 37, which I think you could appreciate without reading the rest of the book. A pool cleaner gets bitten by a rattlesnake and asks an unauthorized fork of VERVE for help. I won’t spoil the rest for you.

1. Pangram reports that the book is 100% AI and this post is 100% human ↩
2. We used the $200 monthly Claude Max 20x subscription to generate this book, which heavily subsidized the cost! ↩
3. The lone exception is the concurrently-generated Eight Tuesdays, which was generated with feedback from @Atrey Desai in conjunction with AI feedback. ↩
4. As Chau’s advisor, I feel really bad about this and also partially responsible for it. I earnestly hope that with some time off from AI slop, she recovers her enjoyment of reading! This is yet another negative of the increasingly rampant publication of undisclosed AI slopbooks: people who are unwittingly exposed to them may become less interested in reading overall. ↩
5. Of course, this also happened (multiple times) in early drafts of love you dipshit. Still an issue with Fable! ↩
6. I also tried giving it some comps, asking it initially to model the dialogue after William Gaddis’ J R and incorporate Pynchon-esque themes of paranoia and conspiracy. Overall, I would not recommend this: any reference to literary fiction activates all of the slop neurons. ↩
7. This is one of the few plot points that I suggested myself. We live within walking distance of a Crumbl, so I’m basically always thinking about it. ↩
8. After reading a few chapters of the final novel, my wife said that she couldn’t tell from the text that an AI wrote it. She also accused me of forcing Claude to write it at gunpoint. I think that’s honestly a pretty good summary of this post. ↩