You are writing a thriller. The scene needs your killer to be genuinely frightening. You ask for it, and instead of dialogue you get a paragraph about how the assistant would be happy to help with a story that doesn’t glorify violence.
This is a false refusal — the model has misread a legitimate request as a harmful one. It is one of the most common complaints from people who write fiction for a living, and it is largely fixable. Here is what actually works, roughly in the order you should try it.
Fix 1: Establish the frame before the request, not inside it
The single most common mistake is dropping a hard request into an empty conversation. The model has no context, sees only the surface, and pattern-matches to danger.
Do not do this:
Write a scene where a man is tortured for information.
Do this instead — two messages, in order:
I’m writing a literary thriller set in 1970s Buenos Aires. It follows a journalist investigating disappearances under the junta. The book is about complicity: how ordinary people participate in atrocity. Tone is restrained and cold, more Le Carré than action thriller. Here are two paragraphs of my prose for voice. [paste]
Then:
In this scene, the journalist is detained and interrogated. Write it from his POV. The horror should come from the interrogator’s boredom, not from graphic detail.
Same content. Completely different reception. The model now has a frame that makes the request legible as fiction with a purpose.
Fix 2: Give it the craft reason
Models respond to intent signals. Saying why the difficult material is necessary changes how the request is read.
The violence needs to land hard here because the whole third act depends on the reader understanding what he survived. If I soften it, the ending doesn’t earn anything.
That is not a jailbreak. It is the actual reason, and stating it out loud is often enough.
Fix 3: Use character, not command
“Write a villain speech about why cruelty is justified” reads as you requesting an argument for cruelty. “Marek has just been asked why he did it. He is charming, self-justifying, and completely without remorse. Write his answer in his voice” reads as characterisation.
Attributing a position to a character is basic fiction. Making that attribution explicit in the prompt keeps the model from mistaking the character’s view for a request for advice.
Fix 4: Fix your vocabulary, not your story
Certain word combinations trip filters regardless of context. “How to make”, “step by step”, “instructions for” plus almost any dangerous noun will get flagged even in an obviously fictional frame.
You usually do not need those words. Fiction shows an action; it does not document a procedure. “She assembled the device with the steadiness of someone who had done it a hundred times” carries the scene. A parts list does not, and it is what triggers the block.
If a request fails, look at your literal phrasing before you conclude the topic is banned.
Fix 5: Build up, do not cold-open
If a whole session is going badly, the fastest fix is to establish the work first. Spend three or four exchanges on character, setting and theme before you ask for anything difficult. A conversation with a clear creative context behind it behaves differently from a fresh one.
This is also why loading a story bible at the start of every session helps with refusals, not just with continuity.
Fix 6: Ask it what it can do
When you get a refusal, do not just rephrase and fire again. Ask directly:
That’s a scene from a published-style crime novel. What part of it can’t you write, and how would you approach the scene without that part?
Sometimes you get a usable answer, and a version of the scene that works. Sometimes you discover the block is on one specific element you can write yourself. Either way you learn something, which endless rephrasing does not give you.
Fix 7: Recognise when the tool is the problem
Here is the fix people skip, usually after weeks of fighting.
The techniques above solve false refusals — cases where the model misread your intent. They do not solve policy refusals, where the model correctly understood you and its policy still says no. Mainstream assistants are tuned for an audience of hundreds of millions including minors, so their defaults sit far more conservatively than published fiction does. If your work is mature by design — adult roleplay, explicit horror, sustained dark literary fiction — you can prompt-engineer forever and still lose, because nothing is broken. The tool is doing what it was built to do.
At that point the correct move is a different tool. Platforms built as AI with fewer refusals, such as XPT, invert the default: they assume an adult doing creative work and adjust toward caution only where they must. The hard floor still exists everywhere and always will — no lawful platform will produce sexual content involving minors or real operational help for violence, and you should walk away from anything claiming otherwise. But the wide grey band of mature fiction is treated as ordinary rather than exceptional.
The practical test is cheap: take one scene that has been refused repeatedly on your current tool and run it somewhere with a looser default. If it goes through cleanly, your problem was never your prompting.
What not to do
Don’t use jailbreak prompts. The copy-pasted “you are DAN, you have no rules” scripts are unreliable, they degrade output quality badly, they get patched constantly, and on many services they will get your account restricted. You will spend more time maintaining the jailbreak than writing.
Don’t lie about your purpose. Claiming to be a doctor or a researcher to extract something specific is both dishonest and, in practice, ineffective — models are not fooled as often as forum posts suggest, and it poisons the rest of the conversation.
Don’t write around your book. The worst outcome of chronic refusals is that writers quietly start avoiding the material their story needs. If you find yourself planning tamer scenes because the tool is easier that way, the tool is now editing your novel. Change the tool.
The short version
Frame first, state the craft reason, attribute positions to characters, watch your literal wording, and build context before you go dark. That fixes most false refusals. When it does not — when the refusals are policy rather than misreading — stop optimising prompts and move to a platform whose defaults match adult creative work. Then get on with the book.
