Analyzing Fiction Writing with AI (Reverse-Analysis Method)

Learning things I didn't know about my writing through teaching Claude to write fiction like my pen name.

You learn the most by trying to explain what you know.

Writing works the same way. But without a student to reflect your teaching back to you, there’s no feedback loop. So I built one. I turned Claude into a student that teaches me back.

Most writers use AI to analyze finished writing. I’ve stumbled on something more effective: get the AI to write first, then show it where it went wrong.

Why Standard Analysis Falls Short

Most AI writing tools lack a north star. Hand an AI a writing sample and it can infer patterns, but it doesn’t understand your intent. The constraints are not clearly defined. It doesn’t know what you’d never say, where your metaphors come from, or the specific vibe of your voice.

The Reverse Workflow

Instead of starting with a writing sample, I built a Writer Profile first, essentially an SOP for a ghostwriter. It includes voice and tone constraints, signature moves, banned words and overused phrases, POV execution rules, and genre-specific tropes (like the obligatory beats in an enemies-to-lovers romance).

Then I run a three-step loop:

1. Voice Test Draft: Claude writes a 500-word scene from scratch based only on the Profile. What it writes is entirely up to the AI. It invents the characters, setting, and story.

2. The Rewrite: I rewrite that scene in my actual voice.

3. The Delta Report: Claude analyzes the differences between its rule-based version and my natural one.

Magic in the Delta

By comparing its attempt to my rewrite, Claude surfaces the invisible elements of my style, the things I couldn’t articulate myself. It updates the Writer Profile with these new nuances and produces another, better draft for the next round.

Because the 500-word story is entirely the AI’s creation, I can play ruthless editor with zero emotional attachment. I have no intent of publishing this as a story. It’s not a precious darling I’ve decided to kill. It’s simply a test.

The more passes we do, the closer Claude gets to writing like me, and the more it reveals about my own craft in each report.

Where Style Meets Genre Rules

Every genre has rules: heat levels, violence thresholds, language norms. AI can handle the structural stuff well enough: plot beats, tropes and other genre conventions.

It struggles with what some people might call the “soul” of the writing.

It includes things like word choice, sentence cadence, and emotional resonance. Those primarily come from the writer, although in reality they’re just a more complex set of patterns that AI may eventually master.

By forcing the AI to try (and fail) to mimic you from a set of rules, then correcting it, the AI picks up on nuances it couldn’t extract from either rules alone, or a writing sample alone.

In the process, you see your own “writer’s soul” (for lack of a better term) reflected in its analysis of where your writing diverges from its draft.

Through this process, you’re not just teaching an AI to write like you. You’re getting feedback on your writing that allows you to finally see your own style clearly enough to define it.

The Reverse Analysis Exercise

Instead of having AI analyze your writing, reverse it and have AI write based on a profile of how you write. Here’s an exercise:

Step 1 - Pen Name (or Writer) Profile: Create a detailed pen name profile explaining how your pen name writes in a particular genre with particular tropes, POV, language choices, and so on.

Hint — have the AI interview you and ask you questions about how you write. Make sure to include genre and trope information so the AI gets the full flavor and context.

Step 2 - The AI Writes: Have Claude (or the AI of your choice) write a 500 word writing sample using the Writer Profile you just created.

Step 3 - Rewrite and Update: Rewrite the sample and give it back to the AI. Ask it to pick up on the differences and recommend updates to the Profile. You might just learn something about your writing. I know I did.

Here’s Claude’s response to my first run through:

I can read several clear directional shifts in your rewrite. Let me break down what I’m seeing:

1. Grounding in the unglamorous and specific. You replaced the slightly poeticized “woodsmoke and old beer” opening with something more visceral and unpleasant — “trying not to gag on the smell of old beer and body odor that seemed to permeate the walls.” The bar in my version was atmospheric-dark. The bar in your version is a gross bar. The black-and-tan is a real drink, not “something dark.” This is a significant signal: Your Pen Name’s world has grit and specificity, not curated moodiness. The setting isn’t aestheticized for the reader’s benefit — it’s experienced the way the character actually experiences it.

2. Character behavior reveals character. The napkin detail is the biggest addition. [The MC] pulls napkins from the dispenser and puts her fingertips on them rather than touch the table. Then “Did anyone ever clean this place?” — that single interior line tells you more about who she is than any description could. She’s practical, slightly fastidious, not someone who romanticizes the situation she’s walking into. My version had her set her hands flat on the table as a power move. Your version has her manage the disgustingness of the table first and then keep her hands visible. The power dynamic is still there, but it’s layered with personality.

3. Action and description interwoven, not sequential. My version had discrete blocks: description paragraph, then action, then description. Your version weaves them — she’s noticing his face while she’s deciding whether to put her hands on the table, she’s tracking the copper smell while she’s checking exit routes. The observation and the doing happen simultaneously, which reads as more naturalistic and more tense.

4. The “ears getting warm” replacement. You replaced “her pulse picked up” (internal, invisible, told) with “her ears were getting warm” (physical, visible, shown) and paired it with a behavioral response — checking exit routes. That’s two things at once: the involuntary physical response (attraction or fear or both — the ambiguity is good) and the trained tactical response (she’s someone who checks exits). My version named the feeling and managed it. Your version shows two competing bodily realities without explaining either one.

5. Slightly more distance in the narration. My version sat very close to [the MC’s] interiority. Your version pulls back just a half-step — we’re still in her POV, but the narration has a bit more behavioral observation and a bit less sensory immersion. “She started to set her hands on the table but then pulled back, his eyes tracking” — that’s almost cinematic, watching her from the outside while still being inside her head.

[…]

I’d adjust a few specs based on your rewrite:

Vocabulary register: Add that Pen Name’s settings are rendered with unglamorous specificity. The world is not aestheticized — bars smell bad, tables are sticky, real brand names and real drinks appear. Grit is authentic, not curated.

Interiority style: The profile says “fragmented, sensory, biased.” Your rewrite says interiority is also practical and behavioral — [the MC]’s interior thoughts are about whether the table is clean and where the exits are, not about her feelings. Emotions are rendered through involuntary physical responses (ears warming), not through interior acknowledgment.

Signature move #1 (physicality as emotional language): Should specify that the physicality includes behavioral choices (napkins, exit-checking, hand placement) alongside involuntary responses. Characters reveal themselves through what they do in a space, not just what their bodies do to them.

Narrative distance: The profile should note that the close third sits at a slightly more behavioral distance than deep-POV — we see the character acting and reacting, not just feeling. It’s close enough to access interiority but pulled back enough to show behavior as the primary channel.

What I Learned

You learn the most by trying to explain what you know, even if the student is an AI. I learned a few things from this, but what stands out is what surprised me.

In reading the report, I noticed that my idea of “third person close” was different from how I imagined, and different from how I originally described it. The way Claude explained that narrative distance element started me thinking, and then it hit me…

It makes sense that I prefer movies and TV over books. Shocking, I know!

Yes, I’m weird that way. I’m very visual, so I want someone to show me their version of the scene as they see it. I’ll figure out what it means and what the characters are feeling myself. Apparently, that’s how I write as well. Maybe screenplays are in my future.

In any case, this is much stronger feedback than blind analysis, and more effective than giving the AI writing samples alone.

AI analysis is all about pattern matching and at the end of the day, writing is just patterns. If that weren’t true, how could you ever improve? If there were no patterns, It would be pointless to analyze writing and grade students in a writing class. It’s all just patterns.

Writing profile + Claude sample + rewrite + analysis.

Try it and let me know how it works for you. What did you learn?

What’s Next?

There are lots of things we can do with this. For example, we could analyze the writing and score it on a points scale for several parameters to see how close your writing comes to meeting requirements.

Things like tone, pace, genre, tropes, tension axes, language, and romantic arcs.

We could create beta reader or editing modules with editing suggestions and tracking. In other words, you can create a workflow where you write, and then get almost instant feedback on how close the writing comes to expectations.