"My ghostwriters give the skeleton. The flesh and blood come from Dumas."
Alexandre Dumas said that in the 1840s, defending himself against critics who called him a literary factory boss. His most prolific collaborator, Auguste Maquet, produced draft after draft -- structurally sound, logically coherent, perfectly readable. Dumas rewrote every sentence. Maquet's solo career never matched those heights. The missing ingredient wasn't craft. It was voice.
Fast forward a hundred and eighty years. The ghostwriter got an upgrade. Faster, cheaper, never complains, available at 3 AM. And the output? Smooth. Correct. Complete.
Also hollow.
Read AI-generated prose and something registers as off. Not wrong, exactly. The grammar holds. The metaphors land within acceptable range. But the rhythm isn't yours, the observations aren't yours, that quality of "nobody else would phrase it this way" -- absent.
This is AI's mediocrity problem. Not a bug. The inevitable result of how language models work. And if Dumas could fix it with Maquet's drafts, the same approach works here.
Why Does AI Output Tend Toward Mediocrity?
Every language model does one thing: predict the next most probable word.
Probable. Not best.
That distinction determines everything. Faced with "describe sadness," the model scans millions of sadness descriptions in its training data and averages them out. "Tears slid down her cheeks" has the highest frequency, so out it comes. "Silver moonlight bathing the earth" -- highest frequency, so there it goes again. These expressions aren't wrong. They're average. The first answer everyone thinks of.
Good writing lives just past that first answer.
Writers take risks. Metaphors nobody has tried. Sentences broken into ungrammatical fragments. Characters laughing when they should be crying. Sometimes these choices fail spectacularly. But they're also where surprise is born. AI is trained to produce the "most likely correct" answer, not the "most likely to make someone sit up straight" answer. Its job is arriving safely. Not delivering impact.
There's a deeper layer.
AI has no memory of experience. Describing "morning coffee," it won't recall a mother grinding beans in the kitchen, won't summon the burning sensation of a fourth Americano during finals week, won't think of that particular morning -- the coffee went cold because a phone call changed everything. These lived experiences seep into the capillaries of prose, making two people write the same subject entirely differently. AI can mimic "personal" tone. But what it extracts is statistical pattern, not anything actually lived.
So it delivers "generic morning coffee." Technically correct. Emotionally vacant.
The Dumas Method: Always Rewrite
In 1844, Dumas hired Auguste Maquet as a ghostwriter. Maquet wrote first drafts based on Dumas's concepts. Then Dumas did one thing -- rewrote every sentence his own way.
Critics mocked him as a "literary factory boss."
He didn't care. "The ghostwriter gives the skeleton. Flesh and blood come from Dumas."
Maquet's drafts were fluid, well-structured, logically clear. But something was always missing when reading them. It was Dumas's rewriting that turned The Three Musketeers and The Count of Monte Cristo into classics. Maquet later attempted independent writing. Never reached the same heights. The difference came down to one word -- voice.
The right attitude toward AI output mirrors Dumas's treatment of Maquet exactly: raw material, not finished product.
Not "editing." Letting someone else's words ignite your own.
An example. Ask AI to describe a sad scene. It produces:
She felt a deep sadness, tears streaming down her cheeks.
Standard "most probable" sadness. Now rewrite it the Dumas way:
She buried her face in the sofa cushion -- the one that still held traces of his cologne -- and discovered she couldn't cry at all.
Same sadness. But the second version has concrete detail (a cologne-scented cushion), surprise (unable to cry), personality. That "couldn't cry at all" is real grief -- pain so acute that tears won't even activate.
In Slima's Writing Studio, press Cmd+ to open Split Window. AI's raw output on the left, your rewrite on the right. Reference the framework and ideas while re-expressing everything in your own rhythm. When you're done, delete the left panel. It served its purpose.
Force AI Out of Its Comfort Zone
AI always defaults to the safest option. Unless explicitly told: I don't want safe.
A typical prompt: "Describe someone waiting for an important phone call." AI writes gripping the phone, pacing back and forth, staring at the screen. Standard nervous behaviors. Try a different approach -- "Describe someone waiting for an important phone call, but ban all common anxiety signals. No gripping the phone, no pacing, no staring at the screen."
Something interesting happens.
AI gets backed into a corner and has to explore territory it would normally avoid. Maybe the character starts dusting the bookshelf. Maybe he counts birds on the power line outside. Maybe she opens the fridge, closes it, opens it again, closes it -- with no memory of what she was looking for.
In the AI Assistant (Cmd+Shift+A), try this prompt template:
Rewrite this description in an "unsafe" way.
Specific requirements:
1. No common metaphors or cliches
2. Give me three versions: one safe, one weird, one dangerous
3. The dangerous version can challenge reader expectations
I'll pick one as inspiration, then rewrite it my own way.
"Three versions" is the key. AI's first instinct is always safe. But when forced to produce "weird" and "dangerous" variants, it searches the edges of its statistical distribution. Those versions might not be perfect -- they might even be absurd. But absurdity hides inspiration. It shows the writer landscapes beyond the first option.
Teach AI Your Voice
Imitation requires samples. Saying "write in my style" is as meaningless as saying "write in a good style" -- AI has no idea what either means.
But give it a passage you're genuinely satisfied with, and the situation changes completely.
The File Tree in Writing Studio makes this radically practical. Create a file called "Style Samples." Collect your best paragraphs -- the ones where rhythm, imagery, and tone feel unmistakably yours. Three to five passages. That's enough.
When AI help is needed, open the AI Assistant and use @ to reference the file:
@Style Samples
This is my writing style. Analyze the characteristics of these paragraphs, then rewrite the following content in a similar style:
[paste content to rewrite]
A more advanced move -- flip the process. Let AI analyze the style itself first. Select sample text and ask:
Analyze this writing's characteristics:
- Sentence length and structure
- Punctuation habits
- Preferred imagery and metaphor types
- Narrative rhythm
Then use these characteristics to rewrite the following passage.
AI imitation will never be perfect. At best, approximate. But "approximate" is vastly better than "no reference at all." And there's an unexpected bonus -- when AI articulates the features of your style, those descriptions often help the writer see their own habits clearly for the first time. Turns out I favor short sentences at paragraph endings. Turns out I gravitate toward food metaphors. Turns out my rhythm runs long-long-short.
These discoveries have value on their own.
AI Is the Diagnostician, Not the Doctor
Switch the dynamic. Don't ask AI to write. Ask it to read.
Select a passage you wrote. In the AI Assistant, ask specific questions: "Where is this description too vague?" "Where does this dialogue not sound like something this character would say?" "Where does this scene's pacing drag?"
Then fix it yourself.
The advantage is fundamental -- the voice stays entirely yours, but you borrow AI's analytical capacity. AI flags "the second sentence is too long, it breaks the rhythm" or "this is telling, not showing." Valuable diagnosis. How to operate? The writer decides.
In Writing Studio, this workflow is seamless. Select text, Cmd+Shift+A summons the AI Assistant, let it analyze problems, then edit directly in the editor. No window switching. No copy-paste. No broken train of thought.
Before editing, press Cmd+Shift+G to create a Snapshot. If the revision doesn't satisfy, return to the previous version any time. This safety net enables aggressive editing -- no tiptoeing, no fear of breaking things.
Structure for AI, Details for You
One final division of labor. The most practical one.
What does AI do well? Build frameworks, ensure logical coherence, generate options. What do humans do well? Unique observations, personal detail, unexpected metaphor, the authentic texture of emotion. Combine both strengths -- that's the most effective collaboration model.
Writing an argument scene. Tell AI: "Build the framework for this argument -- who speaks first, how emotions escalate, how it ends. But don't write specific dialogue." AI delivers the skeleton. The dialogue comes from the writer. Those lines that only these two characters would ever say to each other, loaded with old wounds and private codes -- AI can't produce those.
Designing a room works the same way. "What objects should be in this room to reflect the character's personality? Give me options." AI generates a list. The writer picks and adds details only they would know -- the book on the shelf that's falling apart isn't a classic novel, it's an expired travel magazine, open to page 47 where a circled address marks a guesthouse that never got visited.
The File Tree is built for this workflow. Dedicated note files for AI-generated frameworks and options. In main chapter files, Split Window open with notes on one side. When the writing is done, those framework notes don't need deleting -- Version Control preserves them. Looking back later at how a bare skeleton grew into a full scene is itself a form of learning.
Practice: Rewriting Mediocre to Unique
A complete exercise.
Ask AI to describe "someone returning to their childhood home." It will probably produce something like:
As he pushed open that familiar door, childhood memories flooded back like a tide. Everything here seemed smaller -- the once towering staircase was now just normal height, the once spacious living room now felt cramped. Sunlight filtered through windows, dust floating in the beams. He felt a faint sadness.
Dissect where it's too safe. "Memories flooded back like a tide" -- cliche, used by everyone. "Everything seemed smaller" -- the standard observation every adult makes returning home. "Dust floating in beams" -- used thousands of times in film. "A faint sadness" -- telling the reader the emotion instead of letting them feel it.
Now ask a different question: returning to the old house, what would you actually notice? Not the generalization "everything got smaller." Something hyper-specific. Something only you would register.
Rewrite:
The doorknob was at the wrong height. He remembered reaching up for it -- now it hit his waist. The wall still carried that line -- ten, eleven, twelve -- where his mother had marked his height in pencil. The marks stopped at twelve. That was the year his father left and they moved away. He found himself thinking a strange question: how much had he grown since then?
Concrete detail replaced generalization (doorknob height, not "everything got smaller"). Personal history replaced universal experience (the mother's pencil marks, the father's departure). A strange question replaced "a faint sadness." Readers don't need to be told he's devastated -- the line that stopped at twelve already says everything.
That's the journey from mediocre to distinct. Not magic. Discipline. Every time AI delivers output, ask: where is this too safe? Where can I substitute a detail only I know?
Dumas was once asked a pointed question: "Why not let the ghostwriter finish it and just put your name on it?"
His answer was a single sentence: "Because then it wouldn't be my book. Readers don't buy the story -- stories are everywhere. Readers buy the way I tell the story."
A hundred and eighty years later, that sentence has never been more accurate.
AI can deliver a skeleton. Smooth, correct, complete. But a skeleton isn't a book. What makes the skeleton grow flesh is the detail only this writer would notice, the metaphor only this writer would choose, the twist only this writer would imagine.
AI is this era's Maquet -- diligent, cheap, always on call. But it cannot replace the person sitting there, rewriting every sentence their own way.
This isn't a burden. It's a privilege.