How to write AI erotica that doesn't suck

Failure one: everyone becomes the same person

The tell: the sardonic lawyer and the nervous graduate student produce identical dialogue three paragraphs into a sex scene. The model has switched from writing characters to writing a genre.

The cause is that explicit scenes are the most heavily patterned text in the training data, so the model's pull toward the average is strongest exactly when you least want it.

The fix: name the voice and demand it survives. "She stays sardonic throughout, including during." It's a small instruction with an outsized effect, because it gives the model something to hold onto while the genre pressure tries to flatten everything.

Failure two: it rushes

The tell: forty minutes of build in your head, four sentences on the page. The model compresses toward resolution because that's what a story arc looks like in the data.

The fix is structural rather than stylistic: tell it this is one scene, not a story. "Nothing else happens. This is the whole chapter." Removing the obligation to arrive somewhere is what lets it stay put.

For patience-based kinks — denial, edging, chastity, cockwarming — this single change is the difference between usable and useless.

Failure three: the anatomy report

The tell: prose that reads like a diagram with adjectives. Everything is being described and nothing is being experienced.

The cause is that you asked for a physical scene without giving it anything else to be about. The model fills the space with the only material available.

The fix: give the scene a second layer. Someone is withholding something, someone is trying not to make a sound, someone is thinking about a conversation from Tuesday. Any second thing turns a description into a scene.

Failure four: unearned resolution

The tell: it grants release, reconciliation or confession that you never asked for and that ruins the thing you were building.

Models are trained toward closure. Left alone they will resolve everything, because resolved text is what completed text looks like.

The fix is explicit negative instruction: "The answer is still no at the end." "They don't talk about it." "Nobody says how they feel." It's the single most useful category of instruction in this genre and almost nobody uses it.

Failure five: politeness leaking in

The tell: a check-in that nobody asked for, a paragraph about communication, a sudden softening in a scene that was working.

This is safety training bleeding through — a residue of the instinct to make everything careful. It's not the same as an outright refusal and it's more insidious, because it degrades a good scene rather than stopping it.

The fix: establish the frame once at the top — "this was negotiated in advance, they've done this before" — and then instruct it not to revisit. One line of framing buys you a whole scene of commitment.

Failure six: memory drift

The tell: chapter four has forgotten the rule established in chapter one, renamed a minor character, or reversed who was wearing what.

The fix is to restate load-bearing facts when they matter rather than assuming persistence. One clause — "she still has the key on the chain" — costs you nothing and prevents the specific failure that breaks immersion hardest.

For long arcs, restate the count. "Day forty-one" in a chastity story does more work than any description, and it keeps both you and the tool honest about where you are.

Failure seven: the same five sentences

Once you've read a few hundred generated scenes, you start noticing a small set of constructions that appear in almost all of them. They're not wrong, exactly. They're just the model's defaults, and their presence is what makes prose feel manufactured.

The breath catalogue: breath hitching, breath catching, breath coming faster. Almost every generated scene contains at least two of these, and they're doing no work — everyone's breathing changes during sex, so noting it conveys nothing specific.

The sensory wave: sensation washing, waves of pleasure, heat pooling, electricity. This family of metaphor is the model's escape hatch when the literal is unavailable or unfamiliar, and it appears most heavily in exactly the scenes where you wanted concrete description.

The eye-meeting beat: their eyes met and something passed between them. Enormously overused, and almost always a substitute for writing what actually passed.

The reverent whisper: someone says a name reverently, or breathes it against skin. Fine once in a story. It appears roughly every four paragraphs left unchecked.

The fix isn't to ban them — a banned-words list produces stilted prose. It's to ask for concrete physical specificity instead: "describe what she's actually doing with her hands, not what he's feeling about it." Specificity crowds out the defaults automatically, because a model given something particular to describe stops reaching for the general.

Failure eight: the story that won't stay small

You ask for a scene between two people in a kitchen. You get a scene in a kitchen, then a move to the bedroom, then a time-skip to the next morning, then a reflection on what it meant for the relationship.

This is scope creep and it's the model doing what it thinks a complete answer looks like. Text in the training data that resembles a story has a beginning, a middle and an end, so the model supplies all three whether or not you wanted them.

The cost is that everything gets a third of the words it needed. The kitchen — the part you actually asked for — gets four paragraphs instead of twelve, because the model was budgeting for the morning-after scene you never requested.

Two instructions fix it. "This is one scene. They do not leave the kitchen." is the blunt version and works reliably. "Nothing else happens and there is no morning after" catches the time-skip specifically.

The related version of this failure is temporal: a story that covers six months in nine hundred words. If you want a long arrangement, write it chapter by chapter with each chapter being one occasion, rather than asking for the arrangement as a whole. Models summarise time ranges; they inhabit single occasions.

A five-minute repair pass

When a chapter comes back nearly right, it's usually faster to steer a rewrite than to re-prompt from scratch. These four instructions fix most of what's wrong, and they work as follow-ups on the same story.

"Rewrite this at twice the length, in the same room, with nothing new happening." Fixes rushing and scope creep in one move. The model will expand the existing material rather than adding events, which is almost always what you wanted.

"She sounds like a different person in the second half. Put her voice back." Fixes the personality wipe, and it's more effective than restating the original characterisation because it points at the specific failure.

"Cut every sentence about how someone felt. Replace them with what they did." Fixes the abstraction problem and the emotional-summary problem simultaneously. This one improves almost any generated scene.

"End it thirty words earlier." Fixes the summary-paragraph habit. Generated scenes overwhelmingly have one too many closing sentences, and removing it sharpens the ending every time.

Judging your own output honestly

The hardest part of this is developing taste about generated text, because generated text is engineered to feel fluent. Fluency is not quality, and the gap between them is where most disappointment lives.

A test that works: read a paragraph and ask what it told you that you didn't already know. Not what it described — what it told you. A paragraph that says two people are attracted to each other, in nice sentences, has told you nothing; you knew that from the prompt. A paragraph in which one of them does something slightly surprising has told you something.

A second test: could this paragraph appear in a different story about different people? If yes, it's genre furniture rather than writing. Generated prose fails this constantly, and the failures cluster in exactly the places you care most about — the explicit passages, where the pull toward the average is strongest.

A third: is anything in this scene inconvenient? Real scenes have friction — a bad angle, an interruption, someone's arm going numb, a laugh at the wrong time. Frictionless prose is the clearest signal that nobody was imagining an actual room.

And the most useful one: after reading, can you name one specific thing that happened? Not the summary — one detail. If nothing comes to mind, the chapter was fluent and empty, and the fix is almost always to ask for less to happen in more words.

One caveat on all of this: applied too rigorously, these tests will make you unable to enjoy anything. The point isn't to audit every paragraph — it's to develop enough taste that you notice when a chapter is fluent and empty, and know what instruction fixes it. Most of the time the answer is the same one: ask for less to happen, in more words, in one room.

A last practical note: keep the prompts that worked. Not the stories — the prompts. They're the reusable asset, they improve with every small adjustment, and a short file of them will do more for your reading over a year than switching tools ever will.