Craft Effective Prompts

Last updated: October 14, 2026

What Is a Prompt?

A prompt is the primary tool for instructing an AI model to generate what you envision. Writing effective prompts is key to getting the results you want. It's important to understand that prompting custom-trained models differs significantly from prompting base models (such as Flux, Qwen, Gemini or other foundation models).

The quality of your prompt directly shapes the results, making it an essential skill to master. This guide walks you through actionable steps and practical tips to help you create prompts that bring your vision to life with precision and efficiency.

Example of a fairly long prompt with a base model (Flux 1.1 Pro Ultra)

Finding the “Prompt” Box

You'll find the prompt box (field) in the left inference panel of the generation interface. This same field appears across various Scenario tools, including Image Generation but also Textures, Skyboxes, Canvas (Retouch), Expand (outpainting), Restyle, or even Enhance (upscale).

While this section focuses primarily on writing prompts for image generation (https://app.scenario.com/images/new), similar principles apply to other tools in Scenario.

Prompt Development is an Iterative Approach

Discovering optimal prompt structure, length, and detail level for any model requires an iterative approach. Scenario's interface is specifically designed to facilitate this process, allowing you to:

  • Test different prompts systematically, one at a time

  • Compare results side by side in the generation panel (right)

  • Identify which prompts produce the best results for your specific needs

Begin with simple prompts, analyze the generated results, and refine based on what you observe. A methodical process helps you develop an understanding of what works best for particular models and projects.

Prompting for base models (e.g. Flux, Qwen, Z-Image)

Base models offer great versatility in prompt handling. They can process both brief descriptions and detailed paragraphs, with longer prompts generally producing more detailed images, as shown below:

However, these base models present significant challenges for production work:

  • Detailed prompts become increasingly difficult to “control“ for consistency across multiple consecutive generations

  • Finding and reproducing a specific, existing art style (or a unique character) using only text prompts can be nearly impossible, requiring extensive experimentation, time and computation.

These limitations make base models well-suited for exploration and concept work, but they might not be the tool of choice for generating consistent production assets due to too much randomness.

Prompting for custom-trained models

Custom AI models dramatically improve consistency by learning specific styles or elements, offering significant advantages over base models. With a custom model, the AI is pre-trained to reproduce particular visual characteristics. This means you no longer need to struggle with complex prompts trying to describe every stylistic nuance or character feature.

This results in significantly more predictable outputs while requiring far less prompt engineering. The time saved by not having to endlessly refine prompts to achieve a consistent style can speed up production workflows, enabling rapid iteration and asset creation.

Here are some core principles when working with custom models:

  • You'll typically use shorter, more streamlined prompts that focus primarily on the content you want to generate (scene, subjects), less on the style.

  • The order of the words is important as the model tends to prioritize elements mentioned earlier in the prompt. 

  • For optimal results, structure your prompts similarly to the captions used during the model's training process, as this provides familiar patterns for the AI to work with. Scenario’s “Prompt Spark” feature helps with this (see below)

A Step-by-Step Approach to Crafting Effective Prompts

Step 1. Start Simple, Evaluate, Then Expand

  • Begin by using Prompt Spark to generate a series of initial prompts

  • Gradually add more details to the prompts (about characters, scenes, or style)

  • Stop expanding when prompts become too complex or difficult to control

  • Experiment with both complete sentences and comma-separated terms (testing is key)

Unlike base models, custom models will maintain the style more consistently, so the length of the prompt won't have as much influence as with base models. However, there comes a point when the prompt becomes too long and its influence will "take over" the influence of the custom model:

In the example above, when the prompt becomes overly long (4th example, to the right), the proportions of the generated character may no longer be consistent with the style of the custom model (as shown below):

Step 2. Use Prompt Spark for Faster Results

Prompt Spark is a unique Scenario tool to help you write good prompts (at least as a starting point), seamlessly and quickly. Prompt Spark feeds on your model training dataset captions and pinned images prompts. You'll find four main features in Prompt Spark:

  1. Generate a New Prompt: Creates a new prompt from scratch, influenced by captions and pinned images  (if any pinned images on the model).

  2. Rewrite Your Prompt: Just type a few words, and Rewrite will expand them into a more comprehensive prompt matching the recommended structure

  3. Translate: The "Translate" feature is convenient if you prefer writing prompts in a language other than English. Simply write your ideas in your preferred language, then use "Translate" to instantly convert them to English. You can further refine the translated prompt using the "Rewrite" function.

  4. Image-to-Prompt: You can import or drag and drop an image into the prompt box, and Prompt Spark will automatically generate a descriptive, ready-to-use prompt based on the uploaded image. You can then edit and refine these suggestions as needed.

Prompt Spark utility: proposals for a target model

The standalone Scenario Prompt Spark utility (model_scenario-prompt-spark) takes a target model, an optional idea and up to five context images, then returns text proposals containing a prompt and parameters. This is a separate entry point from the prompt-box actions described above. It does not run the target model or render an image.

Choose the target model, describe the intended result and any exact wording, optionally attach references, then request proposals. Review the prompt, reference fields and settings before running the target model separately. You can leave the idea empty to explore from scratch.

  1. Open Scenario Prompt Spark and choose the target model.

  2. Write the subject, composition, style, output purpose and hard constraints in the Idea field. For exploration without a brief, leave it empty.

  3. Optionally add context images. State whether you want to preserve their subject, composition, wording or layout.

  4. Choose the requested number of results and Fast or Thorough processing.

  5. Generate the proposals, inspect the prompt and parameters, then run the chosen target model separately when ready.

Turn a game-character brief into a proposal

Input idea

Develop a single full-body character concept for a cozy exploration game: a small red-panda botanist wearing an olive field jacket, a canvas satchel, and practical boots, holding one folded botanical map. Give the character a clear silhouette and expressive, curious posture, with warm storybook colors and a simple pale background. Keep the design readable at thumbnail scale and avoid extra characters, text, or a turnaround grid.

targetModelId=model_openai-gpt-image-2; numResults=3; thoroughness=thorough

Proposed prompt

Storybook character concept art, warm painterly cel-shaded style for a cozy exploration game. A single small red-panda botanist shown full-body, standing in a curious adventure-ready pose, head tilted and one paw raised as if about to examine a plant. She wears a worn olive-green field jacket with rolled sleeves, a sturdy canvas satchel slung across her chest, practical brown leather boots, and a folded botanical map tucked under one arm. Clear readable silhouette, expressive bright eyes, soft fluffy tail curling upward. Simple pale cream background, gentle warm storybook color palette, soft ambient lighting. Centered composition, designed to read clearly at thumbnail scale. No text, no extra characters, no turnaround grid, no watermark.

Proposed settings

{
  "numOutputs": 1,
  "width": 1024,
  "height": 1536,
  "quality": "high",
  "background": "auto"
}

Open the original Prompt Spark proposal on Scenario

Keep campaign wording explicit

Input idea

Design a polished vertical poster for a fictional night garden exhibition. The exact headline is "AFTER DARK" and the only supporting line is "Explore the garden at night". Build the composition around luminous botanical shapes, deep navy negative space, and warm amber highlights, with clear typographic hierarchy and generous margins. Keep all requested wording legible and do not add dates, sponsors, logos, or other text.

targetModelId=model_openai-gpt-image-2; numResults=3; thoroughness=thorough

Proposed prompt

Elegant typographically-driven vertical poster for a fictional night garden exhibition. Deep navy background with generous margins and abundant negative space for a premium gallery feel. At the top, the headline "AFTER DARK" rendered large in a refined high-contrast serif typeface, warm pale gold, centered with clear hierarchy. Directly beneath in smaller refined italic type: "Explore the garden at night". The lower composition features luminous botanical silhouettes ,  glowing flowers, curling leaves, and trailing vines ,  rendered in warm amber and pale gold, softly emitting light against the dark navy. Strong contrast between the deep background and luminous plant forms evokes a magical nighttime garden. Text rendered verbatim, crisp and perfectly legible. No extra words, no dates, no sponsors, no logos, no watermark.

Proposed settings

{
  "numOutputs": 1,
  "width": 1024,
  "height": 1536,
  "quality": "high",
  "background": "auto"
}

Open the original Prompt Spark proposal on Scenario

Explore from an empty idea

Input idea: left empty.

targetModelId=model_openai-gpt-image-2; numResults=4; thoroughness=thorough

Proposed prompt

Editorial lifestyle photography, golden-hour sunlight streaming through a large loft window. A sleek walnut-wood desk holding a single matte ceramic coffee mug, an open leather journal with cream pages, and a small potted succulent in a stone pot. Warm amber directional light casts long soft shadows across the grain of the desk. Softly blurred modern loft interior in the background with exposed brick and hints of greenery. Shallow depth of field, slight overhead three-quarter angle, warm filmic color grading, subtle dust motes in the light beam. No people, no text, no watermark.

Proposed settings

{
  "numOutputs": 1,
  "width": 1536,
  "height": 1024,
  "quality": "high",
  "background": "opaque"
}

Open the original Prompt Spark proposal on Scenario

Use an image as layout context

Input idea

Use the supplied reference graphic reading REVEAL with an underline as layout context for a cinematic title card. Preserve the exact word REVEAL, uppercase letter spacing, centered alignment, and horizontal rule beneath it. Explore a refined brushed-silver treatment on a deep charcoal background with subtle directional light. Add no subtitle, credits, dates, logos, or additional wording. Treat the supplied image as a referenceImages input for image editing. It is an opaque reference graphic, not an inpainting mask; do not propose a mask parameter. Include the supplied reference image in the proposed referenceImages array.

targetModelId=model_openai-gpt-image-2; numResults=1; thoroughness=thorough

View the context image

Proposed prompt

Cinematic title card, refined and minimalist. Using the reference image as exact layout and composition guidance, transform the word "REVEAL" in uppercase, centered, preserving its exact letter spacing and the horizontal rule directly beneath it. Apply a brushed-silver metallic finish to both the typography and the underline, with fine directional grain and soft specular highlights catching the light. Replace the background with a deep charcoal field featuring subtle directional lighting that falls gently across the letters from the upper left, adding depth and a premium sheen. Keep the layout minimal and balanced with generous negative space. Preserve the exact text, spelling, alignment, and proportions. No additional text, subtitles, credits, dates, logos, or watermarks.

Proposed settings

{
  "referenceImages": [
    "[Scenario reference image URL; inspect the original proposal]"
  ],
  "numOutputs": 1,
  "width": 1024,
  "height": 1024,
  "quality": "high",
  "background": "auto"
}

Open the original Prompt Spark proposal on Scenario

Prompt Spark utility parameters

targetModelId

Required. Choose the model whose input signature the proposal must follow. All examples in this guide use GPT Image 2; its width, height and reference-image fields appear in the returned proposals.

prompt

Optional Idea text, up to 10,000 characters. Use a clear creative brief with exact wording and exclusions. The campaign example specifies its headline and supporting line. The empty-idea example omits this field.

images

Optional, up to 5 context images. The REVEAL example uses a title graphic as layout guidance. Check that the returned call places an editing reference in the intended target-model field.

numResults

Request 1 to 10 proposals; the default is 4. A request is not a guarantee: the checkpoint runs requested 3, 3 and 4 results and delivered 2, 1 and 3. Count the actual outputs before choosing a set.

thoroughness

Fast or thorough, with thorough as the default. The schema describes thorough as an additional critique/diversification pass. Both modes returned valid proposals in this run; no measured quality or speed advantage is claimed.

Reviewing Prompt Spark proposals

Result counts can be lower than requested. Proposals can introduce unwanted wording or put a reference in the wrong field, so inspect each one. Signed reference URLs may expire and need refreshing before a later run. The examples here were checked against GPT Image 2; they do not establish the same reliability for every target model.

Step 3. Structure Your Prompt Based on Model Type

Adapt your approach depending on the model's purpose:

  • For a style-based model (trained on a consistent art style), you might focus more on describing what you want to see (a scene, an object for ex), while the style should automatically be carried / applied by the model

  • For a character-based model (trained on a specific subject), you might focus on describing poses, expressions, outfit or other settings, while the overall character's appearance is maintained

Proceed step by step, introducing elements within the prompt, or completing details at the end. You can build your prompt progressively, adding details as needed. Include specific colors, mood indicators, and key elements to improve consistency.

  • For example: "A serene beach at sunset" can become "A serene beach at sunset with golden sand, turquoise waves crashing gently, and palm trees swaying in the foreground"

Step 4. “Reinforce” the Style (When Needed)

With some custom models, you may occasionally notice the AI defaulting to another style than the one you aimed for (typically toward realism). This happens when your prompt includes subjects that have strong associations with realistic images in the base model's training data.

  • For example, if you prompt "A view of the Eiffel Tower," the model might generate some realistic outputs, given the base model was likely trained with tens of thousands of realistic Eiffel Tower images. This strong association can occasionally “override“ your custom style. The same issue can happen with famous landmarks, celebrities, or well-known characters.

When this happens, you can fix it by reinforcing the style in your prompt. Simply add style descriptors either at the beginning or end of your prompt. You might add a brief reinforcement like "stylized" or "cartoon style" at the beginning, or include a more detailed style description at the end such as "in a vibrant cartoon style with bold outlines and simplified forms." These style tokens remind the model to maintain your custom style even when generating subjects with strong realistic associations. 

The same approach works with characters – essentially, reinforce some or all of the key character features in your prompts.

You can also use the prompt embedding of your trained model to automatically add the style to your generation. Go to your model’s details tab to find the prompt embedding setting and cllick on Describe style button. This description will automatically applied to all your prompts, helping maintain style consistency without manual prompt engineering. Especially valuable for difficult-to-train styles or for balancing prompts that might pull toward different aesthetics. See our dedicated article here.

Step 5. Consider Using Prompt Embeddings

Prompt embeddings automatically add consistent tokens to your prompts without manual entry:

If you're unsure which prompt embedding to add, you can auto-generate a recommended prompt embedding directly in the model page

Access This Feature Via API