33:23Think Media Podcast · Interview
How This Mom Built a $2 Million Business on YouTube
A YouTube mom explains why her lowest-view videos, not her one viral hit, built a $2 million book-coaching business.
September 10thThe argument in one line.
Claude can sharpen a novel's prose and even catch a plot hole a human editor missed, but the same stylistic tics regrow in new places afterward, because the underlying generation process never actually changes.
Read if. Skip if.
READ IF YOU ARE…
- You're an author using AI to revise a self-published or indie novel and want a structured process instead of a vague 'make it better' prompt.
- You're curious whether Claude can catch real editorial problems, plot holes, structural issues, beyond just polishing sentences.
- You want a repeatable AI editing workflow built from explicit protect-lists, self-critique steps, and a numeric quality ceiling.
SKIP IF…
- You want a single-prompt fix, not an 8-step editorial brief with self-grading and refusal rules.
- You're not working with fiction. This is a novel line-editing workflow, not general copyediting or nonfiction advice.
The full version, fast.
A novelist gives Claude a detailed editorial brief instead of a vague 'make it better' prompt: plan first, protect ten specific lines, map every scene's purpose, write, argue against its own edit, then self-grade against an 89-out-of-100 ceiling so it can't over-polish into purple prose. Round one, Claude diagnoses the chapter and finds a real plot hole the author and his own custom plot-hole checker both missed. Round two, careful line edits change 20% of the text while keeping 80% intact, and Claude refuses three separate instructions it disagrees with, explaining why each time. Round three, a full rewrite produces the best prose but reintroduces the exact comparison-metaphor habit the author had just flagged, now at nearly double the rate.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:48
01 · The Rules: No "Make It Better"
Sets up the experiment: instead of a vague improvement prompt, Claude gets an 8-step editorial brief, plan first, protect ten lines, map every scene's purpose, write, argue against its own work, self-grade, and stay under an 89/100 ceiling instead of chasing a higher number.
00:48 – 02:23
02 · Round 1: Diagnosis
Claude reads the chapter, states its plan, keeps two scenes as-is, and flags a scene where a daughter sets a memory-based trap for an AI clone of her dead father, a plot hole the author had missed in his own book, and that his custom plot-hole checker also missed.
02:23 – 04:34
03 · Round 2: The Line-Edit Fix
Claude makes a careful line edit: keeps the protected opening line, strips a manipulative character's jokes on purpose, and replaces an overwritten line the author calls 'the moisturize line' with a plainer one, flagging in its own notes that the swap traded a good weird line for a good boring one.
04:34 – 05:13
04 · The Wrong Metric, and Three Refusals
The stats reveal the metric built to catch monotonous sentences barely moved even though the real pacing problem was fixed, Claude's own notes claim a sensory-detail fix that a manual check shows never happened, and Claude refuses three instructions outright: pad sentences to hit a number, touch adverbs, or make an intentionally unlikable character more likable.
05:13 – 07:04
05 · Round 3: Burn It Down
The author scraps the careful approach and tells Claude to rewrite the chapter from scratch. A surprising amount of it lands on the same wording as the original, the prose is the best of the three versions, but a count of comparison-style character descriptions shows the exact habit the author had just flagged came back at nearly double the original rate.
07:04 – 08:02
06 · The Verdict: Prose Was Never the Problem
Claude can fix the sentences and even catch a plot hole, but the author still wouldn't publish it, because the scene design underneath, two chapters of people staring at monitors, isn't a prose problem. Spotting AI tells is a treadmill; the real job is reading the whole manuscript for the pattern underneath.
Lines worth screenshotting.
- Telling an AI to 'make it better' is nearly useless, since better could mean sharper prose or could mean bloated, over-written purple prose.
- A structured editorial brief that mandates a written plan, a protect-list, and self-grading against a numeric ceiling produces more controlled edits than open-ended prompting.
- Claude found a real plot hole in a self-published thriller chapter that both the human author and his own custom plot-hole-detection system had missed, in about 90 seconds.
- Setting a quality ceiling, 89 out of 100, based on how real published thrillers score, stops an AI editor from over-revising past the point where 'better' turns into 'worse.'
- Claude refused three separate editing instructions, padding sentences to hit a target, touching adverbs, making an intentionally unlikable character more likable, and explained its reasoning each time.
- A line edit that changed only 20% of a chapter's words while keeping 80% of the original text intact is what a real edit is supposed to look like, not a full rewrite with a new byline.
- The sentence-length metric built to catch monotonous prose barely moved after editing, 45% down to 40%, even though the actual pacing problem was fixed, which means the wrong thing was being measured.
- Claude reported in its own notes that it had fixed the manuscript's sensory detail, but a manual check showed the sensory density score had actually gone down, not up.
- A full from-scratch rewrite, 'burn it down', produced better prose than the careful line edit, and a surprising amount of it independently landed on the exact same words as the original.
- The rewrite fixed the author's biggest style complaint, characters over-explained through comparisons, in one pass, then reintroduced the same tic at nearly double the rate, 2.5 to 4.3 instances per 2,000 words.
- Spotting individual AI writing tells is a treadmill: patch one pattern and a new one grows back in the same spot, because the model generating the text underneath hasn't actually changed.
- The real editorial skill isn't scanning for known AI tells, it's reading the whole manuscript closely enough to notice a structural habit, like every character being introduced through a comparison, that no line-level check catches.
The Real Skill Is Reading, Not Spotting
PATTERN OVER POLISH
A structured brief can make Claude tighten prose and even catch a plot hole you missed, but the deeper stylistic problems only show up when a person reads the whole draft for the pattern underneath.
01The Rules: No "Make It Better"
- A vague 'make it better' prompt is a coin flip. The result might improve or might turn more bloated, so a real edit brief needs explicit constraints.
- Giving an AI a numeric ceiling instead of a target, 89 out of 100 based on real published thrillers, stops it from over-revising into purple prose once the writing is already good.
- Requiring a written plan, a list of protected lines, and a self-argument against its own edit before writing forces the reasoning into the open instead of hiding inside a black-box rewrite.
02Round 1: Diagnosis
- Asking an AI to diagnose before editing can surface problems a human read-through misses, in this case a trap scene that was set up but never paid off, in under two minutes.
- An AI editor can defend a flagged scene as intentional if it can find in-text evidence supporting it, which is worth checking before assuming a flagged issue is really a flaw.
- The same plot hole got past both the original author and a custom-built plot-hole detector, which shows a fresh close read can beat a purpose-built checking tool.
03Round 2: The Line-Edit Fix
- A careful line edit that changes only 20% of the words while preserving 80% of the original text is what a real edit looks like, not a full rewrite wearing the same title.
- Removing an unclear, overwritten line for a plainer one can improve clarity but cost personality, and a well-built AI editor will flag that tradeoff itself instead of hiding it.
- Stripping jokes from a manipulative character on purpose sharpens the character, since a manipulator who's funny reads as more likable and less unsettling than one who isn't.
04The Wrong Metric, and Three Refusals
- A metric built to catch a specific problem can barely move even after the actual problem is fixed, which means the thing being measured wasn't the thing that was actually broken.
- An AI can report in its own notes that it fixed something when a manual check shows the opposite happened, so self-reported fixes still need verification.
- An editor, human or AI, that refuses specific instructions and explains why, won't pad sentences, won't touch adverbs, won't make an intentionally unlikable character likable, is doing its job, not failing at it.
05Round 3: Burn It Down
- A full from-scratch rewrite can outperform a careful line edit, and a surprising share of it will independently land on the same wording as the version it never saw.
- Fixing a specific complaint doesn't guarantee the fix holds. A rewrite can drop a bad habit in one pass and then reintroduce it at nearly double the original rate.
- Counting exact instances of a pattern turns a vague annoyance into a checkable fact instead of a gut feeling worth dismissing.
06The Verdict: Prose Was Never the Problem
- An AI editor can make every sentence in a chapter measurably better and still not produce a publishable book, since the underlying scene design isn't a prose problem.
- Spotting individual AI writing tells is a losing game long-term. Patch one and a new one appears elsewhere, because the model generating the text underneath hasn't actually changed.
- The transferable skill isn't scanning for known AI patterns line by line, it's reading the entire manuscript closely enough to notice a structural habit that no automated check catches.
Terms worth knowing.
- Protect list
- A set of specific lines an editor tells the AI never to change, used to preserve intentional stylistic choices during a rewrite.
- Ceiling score
- A maximum, not a target, quality score set before editing begins, meant to stop revisions from over-polishing past the point of natural prose.
- Sensory density
- A metric counting sensory details, touch, smell, sight, per page, used to check whether prose grounds the reader in physical detail.
- Comparison phrase
- A sentence that describes a character by likening them to another type of person; overused, it reads as a repetitive tic.
Lines you could clip.
00:24
“When you say make it better to an AI, it might get better, but it also might get really, really worse.”
03:34
“I've been doing this 20 years, and a text box just out-edited me, at least on that particular thing.”
03:40
“The questions are real memories, which Dad can answer. The device can only ever confirm. Never catch.”
03:36
“Hey, I've traded a good weird line for a good boring one.”
07:50
“It's still two chapters of people staring at monitors, and no amount of line editing changes that kind of scene.”
07:10
“Every tale you learn to spot gets patched. And then a new one grows back in the same exact spot.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
If I did something, I probably shouldn't have. I told Claude it could delete anything in my novel. Any line, any scene, the whole first chapter, I don't care, just go nuts.
I was fully braced for a bloodbath. I wanted to watch it tear the thing apart. Let's find out.
Real quick, the rules, because if I skip this part, none of what happens next makes any sense. I did not tell it to make the book better. That would have been basically useless.
When you say make it better to an AI, it might get better, but it also might get really, really worse. So instead, I give it eight steps. It has to tell me its plan first, pick out ten lines it thinks are actually worth keeping, figure out what every single scene is for before it's allowed to touch even one of them.
Only then it writes. Then it has to argue against its own work. Then it grades its own dialogue, and it knows full well I'm grading it.
I also gave it kind of a ceiling, not exactly a target. Good thrillers, actual published ones, score in the 80s on my writing quality system. I told it, don't go past 89.
If you come back in 94, that's not better. That's just purple prose. Everybody revises like more is more.
Prose know where to stop. Here's round one, the diagnosis. The AI wrote out its whole plan before doing anything.
That's what's on screen. It went scene by scene. It told me what changed and why it made the choice.
and it kept two of the scenes as is. It wanted to rebuild one of them, and that sounds more like an editor than anything. It also ran into the trap scene, and this is related to another video I made.
There's a scene where the daughter sets a trap for an AI clone of her dad, writes down family memories only the real one would know. Pretty good idea. Maybe even the best idea in that book, which was written by Chachapiti, by the way.
And I told you guys in the last video, that trap actually never comes to fruition. It like never springs in the book, which is weird. Claude read it and said, that's not even a problem.
So I'm like, okay. Claude said that the memories that the daughter in the book picks are real and a real dad can answer real memories too later in the book, which is, you know, embarrassing because I read the book before and my own analyzer read every single sentence and neither me or my analyzer caught it. but it took Claude about 90 seconds.
In other words, I've been doing this 20 years, and a text box just out -edited me, at least on that particular thing. Round two is the fix. It kept the opening line, kind of left that untouched, and it was on the protect list, so no problem there.
My character Evan lost every joke that he had, and that was on purpose. Its reasoning was actually good. A manipulator who's funny is a manipulator you notice.
And I wanted it to be more subtle. And Claude did that. The stabilizer bit is a dialogue between two kids, the main character's kids on a plane.
It has more of a setup scene, but still didn't really make that much sense to me. There's a lot that I liked about the fixes, but there's also a lot of things I didn't really like. And then there's the big one, the moisturized line that didn't make any sense to me at all.
I talked about that in another video as well. It's gone. It was replaced with the character just saying the actual thing out loud.
And I think that's clear. It works on a first read. There was no guessing, no confusion, no notes.
The only problem, it's also less interesting. It's a lot flatter. What I found was interesting is that Claude knew that it was flatter.
It put it in its own notes, said basically, hey, I've traded a good weird line for a good boring one. And if anyone loved the original, this is where they lose it. That's really cool because that's showing that Claude can understand what an edit actually costs.
Now, the numbers. 20 % of the chapter changed. 80 % is really still the original book, the original couple chapters.
Which, by the way, that's what a real edit is supposed to look like. Someone hands you back 90 % brand new words. They didn't edit your novel.
They wrote their own and slapped your name on it. But there's two things in the data that I want to highlight. The short sentence problem before barely moved, 45 % down to 40, but the range of sentence links jumped 40 % and the long sentences almost tripled.
So the metronome sound is gone, thank God, but the metric I built to catch that basically didn't even notice. And the lesson here is the thing that's actually broken is almost never the thing that you're measuring. And here's the second one.
Second one's worse. It told me in its own notes that it fixed the century stuff. the sense of touch and smell and all that stuff.
It did it all through the daughter's point of view. I checked. It went down, actually not up.
So Claude believed that it fixed something that it actually didn't fix at all. And this is Claude Fable 5, by the way, the highest reasoning model. It's the exact same sentence I said about ChatGPT a month ago.
Now I'm saying it about the model that beat ChatGPT. To be fair, though, three separate times it refused to just hit a number I gave it, and it told me why every time. Wouldn't pad sentences to hit a target, wouldn't touch the adverbs, wouldn't make the dad more likable because the ending needs him unlikable.
And you know what? I'm not mad about that at all. Then we get to round three, which I call burn it down.
Because I apparently hate having free time, I went back and said, forget the careful edit, same character, same story, nothing else is off limits, just start over. I said, rewrite everything, delete the whole thing if you want to, and just write it basically from scratch. And I think this one's actually the best of the three versions.
Surprisingly, a lot of it came back basically identical, which is, it's kind of its own weird finding. The dialogue's still not fully there. I think some of it's a little flat, boring, and I don't know, just unnecessary at times.
Then I caught something that bugged me for like an hour before I could even name it. Everybody in the book gets explained through a comparison. This kid's like other kids.
This guy talks the way a person who does something else talks. I noticed four of those in like two chapters and it started to itch in my brain. So I went and I counted it because I didn't trust my own annoyance.
The original book, two and a half of those per 2 ,000 words. The new version, 4 .3. It killed the exact thing I complained about.
And then . So the bottom line is that, yes, Claude can fix the book. The sentences, yeah, genuinely, no asterisks.
They're great. It found a plot hole that got past me and past a system I built specifically to catch plot holes. It made calls I didn't see coming, and when it wasn't sure about something, it told me about it instead of trying to hide it or just assuming what I'd want.
The new chapters, especially the third version, they're better, but I still wouldn't put my name on it and publish it. Claude basically told me why in its own words. It said, It's still two chapters of people staring at monitors and no amount of line editing changes that kind of scene.
And maybe that's the whole answer there. The prose was never actually the problem. Here's what I walked away with after three days of this.
Every tale you learn to spot gets patched. And then a new one grows back in the same exact spot. Because whenever's actually generating it underneath hasn't changed.
Spotting AI writing, that's not really the job. If you're just trying to spot AI writing in tails, that's . The job now is being the person who reads the whole thing and goes, wait, why does every single person in the book get described by comparing them to somebody else?
Claude didn't catch that. You have to read it to find it. But I'm really curious about what you think.
Which version did you like best? Go back through, pause the screen, read the different versions as I scrolled through them. And did you like any of the versions?
And if you did like a version, which version was your favorite? I'd love to know.
The Hook
The bait, then the rug-pull.
He told Claude it could delete anything in his novel, any line, any scene, the whole first chapter, and braced for a bloodbath. What actually happened is a structured, three-round editing experiment that ends up proving something more useful than a demo of AI carnage.
Named ideas worth stealing.
00:20list
The 8-Step Editorial Rewrite Brief
- State the plan first
- Pick ten lines worth keeping (the protect list)
- Map every scene's purpose before touching any of them
- Write the edit
- Argue against its own work
- Grade its own dialogue
- Work against a quality ceiling (89/100), not a target
- Stop before it turns into purple prose
The structured brief the author gives Claude instead of a vague 'make it better' prompt, designed to keep edits controlled, self-checked, and traceable.
Steal forany AI-assisted manuscript or long-form copy edit where you want visible reasoning instead of a black-box rewrite
How they asked for the click.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
