llm.md

Examples of Good Prompts for Creative Writing

Johnny

Why Generic Ideas Fail as AI Prompts

If you search for “creative writing prompts,” you will find thousands of lists suggesting ideas like “write about a vampire who is afraid of the dark” or “two strangers meet on a train.” These are excellent exercises for human brainstorming, but they fail spectacularly when fed directly into an LLM (Large Language Model).

To understand why, we need to distinguish between a traditional writing prompt and an AI prompt. A writing prompt is merely a seed—an idea meant to spark human imagination. An AI prompt, however, is a set of instructions. It dictates not just what to write, but how to write it. When you hand an LLM a simple idea without structural guidance, the model defaults to its training data’s most statistically probable path. That path almost always leads to cliché, predictable outputs filled with melodramatic dialogue and overused tropes.

The secret to unlocking high-quality fiction from AI lies in constraints. While humans sometimes feel stifled by rules, LLMs thrive on them. Constraints act as creativity boosters for AI because they force the model away from generic averages and into specific, interesting linguistic territories. By defining boundaries—what the story should include, what style it must adopt, and what it must absolutely avoid—you transform a vague request into a precision tool for storytelling.

The Anatomy of a High-Performing Creative Prompt

A well-engineered creative prompt isn’t a single sentence; it is a structured framework. To consistently generate compelling fiction, your prompts should incorporate five core components:

  1. Role: Who is the AI acting as? (e.g., “You are an award-winning literary fiction author.”)
  2. Context: What is the background of the scene or story?
  3. Task: What exactly do you want the AI to produce?
  4. Format: How should the output be structured? (e.g., prose, screenplay format, bullet points).
  5. Tone: What is the emotional resonance or stylistic flavor?

Consider the difference between a poorly constructed request and a highly engineered one:

ComponentBad PromptGood Prompt
Input”Write a scene where two people argue in a diner.""Act as a gritty noir novelist. Write a 500-word scene in first-person past tense where a detective confronts his informant in a late-night diner. The tone should be tense and cynical. Focus heavily on the sensory details of the diner.”
ResultGeneric dialogue, standard setting descriptions, no distinct voice.Stylized prose, clear narrative perspective, atmospheric tension, and specific pacing.

Notice how the good prompt explicitly defines the narrative voice and tense. LLMs frequently default to third-person omniscient or present tense if left unspecified, which can instantly ruin the mood of your piece. Dictating perspective and tense upfront ensures the foundational architecture of your prose aligns with your vision.

Example 1: Character Development & Backstory

Generating complex characters requires more than asking for a biography. You need to engineer a prompt that forces the AI to explore contradictions, flaws, and hidden depths. Using variables allows you to reuse this template across different projects.

Here is a copy-paste example you can adapt:

Prompt: “You are an expert character developer for psychological thrillers. I need you to create a detailed character profile based on the following variables:

  • [Archetype]: The Reluctant Mentor
  • [Secret]: They caused the accident that paralyzed their former protégé.
  • [Conflict]: They are being blackmailed by someone who knows the truth.

Write a 600-word backstory exploring this character’s internal motivations and deep-seated flaws. Do not summarize their life chronologically. Instead, reveal their history through their current anxieties and coping mechanisms.

To guide your style, here is an example of the desired voice: ‘He didn’t drink to forget; he drank to blur the edges of the memories so they wouldn’t cut him when he shifted in his sleep.’ Match this level of introspective, slightly dark prose.”

This prompt utilizes few-shot prompting—a technique where you provide the AI with one or more examples of the desired output style within the prompt itself. By including that single sample sentence at the end, you calibrate the LLM’s vocabulary and rhythm, ensuring the resulting narrative voice matches your expectations rather than defaulting to encyclopedic blandness.

Example 2: Scene Construction with Sensory Details

One of the most common complaints about AI-generated fiction is that it reads like a sterile summary. The model tends to rely heavily on visual descriptions while ignoring the rest of the human sensory experience. To fix this, you must explicitly demand non-visual sensory inputs.

Furthermore, achieving the golden rule of fiction—“show, don’t tell”—requires the use of negative constraints. Negative constraints tell the AI what not to do, which is often more powerful than telling it what to do.

Prompt: “Write a 400-word scene of a protagonist walking into their childhood home after ten years away. Use third-person limited perspective and past tense.

Sensory Requirements: You must include specific details related to smell, texture, and sound. Rely minimally on visual descriptions.

Negative Constraints:

  • Do not use emotion words (e.g., ‘sad’, ‘nostalgic’, ‘scared’, ‘happy’). Show the character’s emotional state entirely through their physical reactions and interactions with the environment.
  • Do not use more than three adjectives in the entire scene. Rely instead on strong, active verbs to carry the descriptive weight.

Maintain a melancholic, slow-paced tone.”

By restricting adjectives, you force the LLM to dig deeper into its vocabulary for stronger verbs, naturally elevating the prose. By banning emotion words, you enforce “show, don’t tell” mechanics programmatically, preventing the AI from lazily summarizing the character’s feelings.

Example 3: Plot Twist Generation & Outlining

Asking an LLM to “give me a plot twist” usually results in predictable clichés (e.g., “the narrator was dead all along”). Generating genuinely surprising narrative turns requires prompt chaining—a method where you break a complex task into a sequence of smaller, dependent prompts.

Step 1: The Outline

“Generate a 7-beat outline for a sci-fi mystery story where a botanist discovers a plant that mimics human speech. Keep the beats brief.”

Step 2: The Critique

“Review the outline you just generated. Identify the weakest, most predictable plot point. Explain why a seasoned reader would see it coming.”

Step 3: The Subversion

“Rewrite that specific plot beat to completely subvert the trope. Avoid supernatural explanations. Ground the twist in hard science or psychological manipulation. Provide three distinct options for this new twist.”

This multi-step approach forces the AI to evaluate its own work before generating alternatives, yielding far more sophisticated results.

For writers who use outlining software like Scrivener or Notion, you can also instruct the AI to format its output structurally:

“Output the final 7-beat outline strictly as a JSON array of objects, where each object contains ‘beat_number’, ‘plot_point’, and ‘tension_level’ keys. Do not include any conversational text outside the JSON block.”

This allows you to seamlessly import your AI-generated plot beats directly into your preferred writing tools.

Advanced Technique: Iterative Refinement

The first draft generated by an LLM is rarely your final product. The true power of AI in creative writing emerges during iterative refinement, where you use follow-up prompts to sculpt the raw text.

If a scene feels sluggish, you can target the pacing directly:

“Rewrite the second paragraph to make it more suspenseful. Shorten the average sentence length to under eight words. Remove all internal monologue and focus only on immediate physical actions.”

If the dialogue sounds robotic or overly formal, you can adjust the naturalness:

“Rewrite this conversation. Make the dialogue more natural by adding interruptions, trailing thoughts, and colloquialisms. Ensure neither character speaks in complete, grammatically perfect sentences.”

Another highly effective advanced technique is role-play prompting for critique. Instead of asking the AI to rewrite immediately, ask it to wear an editor’s hat:

“Act as a ruthless developmental editor specializing in thriller fiction. Read the scene above. Critique it specifically regarding pacing and dialogue realism. Point out exactly where the tension drops and suggest actionable fixes. Do not rewrite the scene yet; just provide the critique.”

This allows you to maintain creative control, using the AI’s analytical capabilities to spot weaknesses you might have missed, before deciding how to implement the fixes yourself.

Common Mistakes in Creative Writing Prompts

Even with the right templates, writers frequently stumble into pitfalls that degrade the quality of their AI-assisted fiction.

Over-constraining the AI While constraints are necessary, stacking too many rigid rules can suffocate the prose. If you dictate every adjective, ban all metaphors, and restrict sentence structure too tightly, the LLM will produce robotic, stilted text. Leave room for the model to operate creatively within the boundaries you set. Balance strict negative constraints with open-ended stylistic goals.

Ignoring the system prompt context Many platforms allow you to set system prompts—overarching instructions that govern the AI’s behavior across an entire conversation. If your system prompt tells the AI to be a “helpful, polite assistant,” it will bleed that cheerful, customer-service tone into your dark fantasy novel. Always ensure your system prompts align with your creative goals, or explicitly override them in your user prompts.

Failing to specify the target audience or reading level An LLM does not inherently know if you are writing a middle-grade adventure or a dense, literary fiction piece for adults. Failing to define your audience results in a generic, middle-of-the-road reading level that satisfies no one. Always include parameters like “Target audience: Young Adult” or “Reading level: College-level literary fiction” to calibrate the complexity of the vocabulary and thematic depth.

Mastering AI for fiction isn’t about replacing your imagination; it’s about learning to communicate your vision with architectural precision. By moving beyond simple idea generation and embracing structural control, you turn a language model into a highly capable co-author.

Download our free Creative Writing Prompt Template Pack or try the PromptCraft Structured Prompt Builder to refine your next story idea.


Related Reading:

Refine Your AI Prompts Automatically

Put the prompt engineering concepts in this guide to work. Use PromptCraft to instantly rewrite, structure, and optimize your prompts.