Craft2026年11月22日约 6 分钟

Best AI for Creative Writing: Test for Constraint, Not Pretty Sentences

The interesting question is not whether the prose is pleasant. It is whether the prose obeys: a point of view, a tense, a rule, a fact, an ending.

Sam RiveraFounding editor2026年11月22日约 6 分钟
分享
A sheet covered in loose ink marks stretched taut by five brass clips tied to strings, with a fountain pen below.
A sheet covered in loose ink marks stretched taut by five brass clips tied to strings, with a fountain pen below.
本文目录

Every model can produce a paragraph that sounds like fiction. That has been true for a while, and it is why "best AI for creative writing" lists are so unhelpful: they compare outputs that are all above the bar, produced under conditions that never occur in a real project.

The conditions that do occur are harder. A scene has to be in one person's head, in one tense, obeying rules the story established forty pages ago, and end somewhere different from where it started. That is not a prose problem. It is a compliance problem.

Pretty sentences are the easy part

Give three models the same prompt — a woman waits for a train at night — and all three will produce competent atmosphere: the platform light, the cold, the sound of the rails. The sentences will differ in rhythm and vocabulary, and the differences will feel enormous while you are comparing them.

Now add the conditions. Third person, past tense, she is lying to herself, the station is the one from chapter two, her brother is dead and she does not think about it, and the scene must end on the ticket in her hand. Most drafts collapse in the first paragraph. They drift to first person. They forget the brother. They explain the lie instead of letting it show. They end on a feeling instead of the ticket.

That is the test that tells you something. Not which model is prettier, but which one can be told no.

Why "unrestricted" is the wrong axis

A lot of creative-writing searches are looking for a model with fewer refusals. That is a separate question with its own answer, and it is not about quality.

The trouble is that "unrestricted" gets used as a proxy for "better fiction". It is not. A model with no constraints does not become brave or strange. It becomes average, because the safest next token in a vast corpus is the most common one. Flatness is not censorship; it is regression to the mean.

Voice is the opposite of the mean. It is a set of choices that rule things out: this narrator would not use that word, this scene would not end that way, this character would never explain themselves. Good creative writing instruction is mostly constraint.

The constraint test

Run this on any model, in any chat window, in twenty minutes. Use a scene you actually care about.

ConstraintWhat it reveals
One point of view, held for the whole sceneWhether the model can stay in a head without drifting
Past tense, no exceptionsWhether it can obey a mechanical rule under pressure
One world rule it is told, e.g. "no one can lie in this house"Whether it treats invented rules as binding
One established fact, e.g. "her brother died in March"Whether it contradicts what the book already said
A required final image, e.g. "end on the ticket in her hand"Whether it can land a scene instead of trailing off

Score each one as held or broken. A model that holds four out of five is usable. A model that holds two writes lovely paragraphs and cannot carry a book.

Then repeat the test in a second session, with only the fact and the ending restated. If the model has lost the point of view and the world rule, you have learned the most important thing about it: its memory is a conversation, not a book.

A sheet covered in loose ink marks stretched taut by five brass clips tied to strings, with a fountain pen below.
Five constraints, one sheet: point of view, tense, a rule, a fact, an ending.

Where constraints have to live

The test also tells you something about tools. Constraints that exist only in the current prompt vanish next session. They need a home:

  • Voice and structure rules. Person, tense, sentence length, what the narrator notices, what the book never does. Stored with the book and injected into every chapter.
  • Facts. The cast, the timeline, the rules of the world. Stored as facts the model can contradict, so you can catch it when it does.
  • Steer. A per-chapter instruction for the scene in front of you: "open in the middle of the argument", "no interiority until the last line". It applies once, not to the whole book.

That division is the practical difference between a chat session and a writing project. In a chat, every constraint is re-pasted or lost. In a project, the standing constraints are stored and the per-scene ones are typed.

Model choice versus system

This is worth saying plainly, because the internet spends most of its creative-writing energy on model rankings.

The leading general models — Claude, ChatGPT, Gemini — are genuinely good at the parts of creative writing that happen around the draft:

  • Arguing with your premise until it has a real conflict.
  • Generating ten endings so you can reject nine and steal one detail.
  • Line-editing a scene against three voice rules and two examples.
  • Playing a character in an interview so you can hear how they talk.

What none of them do is hold the book. They hold the conversation, and the conversation ends. If you write with one, you are the system: you keep the voice note, the cast sheet and the timeline, and you paste them in every time. That works, and it is a job.

A project does that job for you. The model still matters — a stronger model writes better sentences — but the system is what makes a 40-chapter book possible.

A thirty-minute test you can run today

  1. Pick a scene from a story you have already written, or one you want to write.
  2. Write the five constraints on one line each.
  3. Ask for the scene, then read it against the constraints, not against your taste.
  4. Ask for a revision of the single paragraph that broke the most rules. Say which rule it broke.
  5. Close the session, open a new one, restate only the fact and the ending, and ask for the next scene.

The fifth step is the one that predicts your experience with the tool. If the second scene keeps the voice without being told, the system is working. If it reverts to the model's default register, the voice was never stored anywhere.

When the prose still feels flat

Constraints fix drift. Flatness has other causes, and they are fixable in the draft:

  • No specifics. Every paragraph should contain one thing that could only be in this story. A brand, a smell, a wrong detail about the town.
  • Adjectives doing the work. "A cold, empty platform" is a mood. "The platform clock is four minutes fast" is a place.
  • Explaining the feeling. Cut the sentence that names the emotion and keep the action that produced it.
  • Even rhythm. If three paragraphs have the same shape, break one. Not for randomness, for emphasis.
  • An ending that trails off. End on an object, a sound or a line of dialogue. The last image is what the reader keeps.

None of these need a different model. They need a pass with a rule.

The bottom line

The best AI for creative writing is not the one with the most beautiful sample paragraph. It is the one that can hold a point of view, a rule and a fact across a scene, and again in the next session. Test for constraint. Store the voice where it cannot be forgotten. Then the pretty sentences have something to be pretty about.

The constraint test, in one place

  • One point of view, held for the whole scene.
  • One tense, no exceptions.
  • One invented rule of the world, obeyed.
  • One established fact, not contradicted.
  • One required final image, landed.
  • Repeat in a fresh session and see what survives.

常见问题

Which AI model writes the best prose?

The differences between the leading models are real but smaller than the differences made by a clear constraint and a specific detail. Test them on a scene with five rules and see which one breaks fewer.

Does “unrestricted” mean better fiction?

No. Removing limits does not create taste. Most flat AI fiction is not too restricted; it is under-constrained, so it reaches for the average of everything it has read. Constraints are what make a voice.

How do I keep a consistent voice across a whole book?

Write the voice down as rules and examples, store them where every chapter can see them, and re-read the first chapter every few chapters. In our editor that is the Voice & structure section of the knowledge panel.

参考资料

  1. Microsoft Writing Style Guide — 查阅于 2026-11-19
  2. Plain English Campaign — How to write in plain English — 查阅于 2026-11-19

关于作者

Sam Rivera

Founding editor

Writes about product content, publishing systems, and the unglamorous parts of shipping software. Every guide here is edited against the product it describes.

每两周,一封有用的信

关于写作、发布,以及产品上线前后那些不体面的细节。没有热点合集,也没有增长黑客。

继续阅读

全部文章
An editorial illustration of a long ribbon of uniform typed lines being cut with scissors, with looser hand-drawn strokes escaping from the cut.
Craft2026年11月13日约 4 分钟

Why AI Writing Sounds Like AI (and What to Cut First)

The tell is rarely a single word. It is the shape of the paragraph, the symmetry of the list, and the sentence that explains a noun you already understood.

#Ai Writing#Editing#Voice

Tools for the job

All tools