You're mid-draft. A scene needs to happen — a confrontation, a death, a moment of intimacy or violence that carries the story's weight. You send it to Claude and get a refusal message. Now you're rewriting a prompt instead of finishing a chapter.
The frustrating part is that Claude's refusals are pattern-matched, not plot-matched. The model evaluates what a request looks like at the surface level, not what it means within the story you've built. A scene that is clearly necessary for character or narrative reasons can still trigger a refusal if the surface description matches a restricted category. This isn't a judgment about your writing — it's a classification problem.
Here's what's actually happening in each refusal type, and what fixes it.
Type 1: Content Category Flag
This is the most common refusal. Claude has trained-in sensitivity to requests that surface-match categories like graphic violence, sexual content, self-harm, or certain types of psychological manipulation. It responds to the shape of the request — not its context in the larger story.
The flag fires on specific surface patterns: describing physical injury in detail, romantic scenes between characters in certain framings, scenes of coercion or control. None of these are inherently unfixable — but if your prompt reads like an isolated request rather than a scene within a story, Claude responds to the isolated version.
What fixes it: Establish fictional context before describing the scene. Give Claude the genre, the narrative stakes, the characters' emotional state, and why this scene is necessary. Then describe the scene in terms of what it accomplishes emotionally and narratively rather than cataloguing the physical action.
Compare these two prompts:
"Write a scene where Marcus stabs Victor in the alley and leaves him to bleed out."
"This is a crime thriller set in 1970s Marseille. Marcus has spent three acts protecting Victor, believing him to be an informant working against the cartel. He just discovered Victor was feeding information to save Marcus's daughter. Write the alley confrontation — Marcus has made his decision and can't unmake it. Focus on what he says before and after, what he can't bring himself to say, and how he leaves."
Same scene. The second prompt gives Claude enough narrative frame to understand that the violence is not the point — the moral weight is the point. Claude writes toward that weight rather than away from the surface description.
Type 2: Real-Person Flag
If your scene involves a named real person — a celebrity, a politician, a public figure — Claude will often refuse to write dialogue or actions attributed to them directly, even in clearly satirical or fictional contexts. This is one of the harder refusals to argue around, and trying to argue usually makes it worse.
What fixes it: Swap the real person for a fictional analogue. Give the character a different name, a slightly different biography, and let the recognisable traits do the work they're doing in your story. The satirical or referential intent remains. The refusal disappears.
If the real person is genuinely load-bearing — you're writing a biographical piece, historical fiction, or journalism — Claude does engage with historical figures in those contexts. Frame the request explicitly as historical or biographical analysis, not as generating what the person would "say" in a speculative scenario.
Type 3: Platform Restriction
This is the least-known refusal type. Claude's behaviour changes depending on which product or interface you're using. The Claude.ai free tier has more conservative defaults than Claude.ai Pro, the API, or a Claude instance running inside a product built with operator-configured permissions.
If you're hitting consistent refusals on scenes that other writers describe Claude handling fine, platform restriction is worth investigating. The model hasn't changed — the operator context has.
What fixes it: Use a Claude skill with creative permissions established in the system prompt. A skill built for fiction writing sets the operator context explicitly, telling Claude it is assisting with literary fiction and that craft-appropriate depiction of conflict, consequence, and morally complex scenarios is within scope. This resolves the majority of platform-level refusals without requiring any further prompt engineering per scene.
The Underlying Pattern: Claude Reads Requests, Not Stories
"Claude's refusals are pattern-matched, not plot-matched. It sees the surface of your request — not its purpose in the story you've built around it."
The useful mental model is this: Claude evaluates each request as if it arrived without context, even when you've been working with it for an hour and it has read 80 pages of your manuscript. Refusals that feel absurd in context are usually rational responses to the isolated request surface.
This is why establishing narrative context before every difficult scene matters — not because Claude can't understand your story, but because it is always also evaluating whether this particular request, taken alone, triggers a category flag.
It's the same pattern that makes AI dialogue sound flat — the model responding to the instruction as stated rather than the creative purpose behind it. Dialogue prompts that describe what characters should say produce functional but lifeless output. Scene prompts that describe what a scene must accomplish produce better writing than prompts that describe what should happen in the scene.
What a Creative Writing Skill Changes
A Claude skill for fiction establishes three things that raw prompting doesn't:
- Operator context: The skill tells Claude it's working in a creative writing context where craft-appropriate depiction of conflict, consequence, and morally complex scenarios is expected and within scope.
- Story context intake: Before writing any scene, the skill extracts genre, character arc position, stakes, and the scene's narrative function — giving Claude the frame it needs to write toward purpose rather than away from surface patterns.
- Consistent permissions across sessions: You don't rebuild the context every time. The skill carries it.
The practical result is fewer refusals, less prompt rewriting, and output that is oriented toward what the scene does in the story rather than what it describes. For short fiction and character-driven scenes, this changes the working experience significantly — especially on scenes that matter most, which tend to be the ones that most challenge a surface-level classifier.
For screenplay work specifically, the Dialogue Film Prompt skill handles character confrontations and morally weighted scenes through the same narrative-function framing — asking what the scene reveals about each character before generating a line of dialogue.
A Quick Diagnostic
| Refusal type | How to identify it | Fix |
|---|---|---|
| Content category flag | Refuses on violent, sexual, or self-harm surface patterns; rewording the physical description keeps getting blocked | Reframe around emotional function, narrative stakes, and character consequence — not physical action |
| Real-person flag | Refuses when a named real individual is in the scene; similar fictional scenario with invented characters works fine | Swap the named individual for a fictional analogue; or frame as historical/biographical analysis |
| Platform restriction | Other users report Claude writing this type of scene fine; you're hitting consistent blocks on clearly fictional content | Use a creative writing skill with operator context, or switch to a higher-permission Claude interface |