Image Tools

Image Tools

Use this skill when the user asks to create, draw, generate, design, edit, modify, or enhance images.

Loading this skill activates the generate_image and edit_image tools.

IMPORTANT: After loading this skill, you MUST call either generate_image or edit_image tool to perform image operations. Do NOT call image_tools directly - it is NOT a tool name.

Tool: generate_image (text-to-image)

Generate images from scratch using only a text prompt. No reference image is used.

When to Use

• User asks to create a brand new image with no reference image.
• User asks to "regenerate" an image with a refined prompt (starting over, not editing).

Parameters

Example

generate_image(
  prompt="A serene Japanese garden with cherry blossoms, koi pond, and a wooden bridge, watercolor painting style, soft morning light filtering through branches",
  aspect_ratio="16:9",
  n=1
)

The tool returns image URLs in markdown format. Include them in your response.

Tool: edit_image (image-to-image)

Edit or generate an image using one or more reference images as the visual basis. The reference images guide the output — the AI sees them and builds upon them.

When to Use

• Reference-based generation: User provides a reference image and asks to create new content based on it (e.g. "use this character to generate a three-view sheet", "put this person in a fantasy world").
• Character consistency: User wants to maintain a character's appearance across different scenes or poses.
• Style transfer: Transform an image into a different art style while preserving the content.
• Inpainting / Partial edit: Modify a specific region of an image while keeping everything else unchanged (e.g. "change the sofa color", "remove the person in the background").
• Multi-image composition: Combine elements from multiple reference images into a new scene (e.g. "put the dress from image 1 on the person in image 2").
• High-fidelity preservation: Preserve critical details (face, logo, text) while making other changes.

Key decision rule: Whenever a reference image exists (from canvas, chat history, or user upload) and the user wants the output to visually relate to it, use edit_image. Only use generate_image when creating from pure text with no visual reference.

Parameters

* Either image_url or image_urls must be provided. image_urls takes precedence.

Examples

Single-image reference (character in new scene):

edit_image(
  image_url="/api/media/character-ref.png",
  prompt="The same woman from the reference image, now standing in an ancient Chinese xianxia world on a floating mountain path, wearing flowing white immortal robes with jade accessories, cinematic golden-hour lighting, volumetric fog, 8k ultra-detailed",
  aspect_ratio="16:9"
)

Multi-image composition (combine elements):

edit_image(
  image_urls=[
    "/api/media/character-sheet-1.png",
    "/api/media/character-sheet-2.png"
  ],
  prompt="An epic battle scene between the two characters from the reference images. Character 1 (from first image) leaping with a light sword overhead strike, Character 2 (from second image) blocking with dual axes, energy collision sparks. Rain-soaked cyberpunk street at night, neon reflections on wet ground, low-angle cinematic shot, 8k",
  aspect_ratio="16:9"
)

Inpainting (partial edit):

edit_image(
  image_url="/api/media/living-room.png",
  prompt="Using the provided image, change only the blue sofa to a vintage brown leather chesterfield sofa. Keep everything else exactly the same, preserving the original lighting, shadows, and composition"
)

Style transfer:

edit_image(
  image_url="/api/media/city-photo.jpg",
  prompt="Transform this photograph into the artistic style of Vincent van Gogh's Starry Night. Preserve the original composition of buildings and streets, but render all elements with swirling impasto brushstrokes and a dramatic palette of deep blues and bright yellows",
  quality="hd"
)

The tool returns the edited/generated image URL in markdown format.

Decision Guide

Prompt Writing Guide

Core principle: Describe the scene narratively, don't just list keywords.

The model's strength is language understanding. A descriptive paragraph always produces better results than a keyword list.

Structure Template

Build prompts in this order:

1. Subject — Who/what is the main focus?
2. Action/Pose — What are they doing?
3. Setting/Background — Where is this happening?
4. Style — What visual style? (photorealistic, anime, oil painting, pencil sketch, etc.)
5. Technical details — Lighting, camera angle, resolution.

Good prompt:

"A weathered old fisherman mending nets on a wooden dock at golden hour, Mediterranean fishing village in the background with whitewashed buildings and blue shutters, oil painting style with visible brushstrokes, warm amber lighting, eye-level composition"

Bad prompt (keyword soup):

"fisherman, dock, sunset, Mediterranean, oil painting, detailed, 8k"

Style Keywords Reference

Aspect Ratio Guide

Edit Prompt Patterns

When using edit_image, use these proven prompt patterns for best results:

Pattern 1: Inpainting (Partial Edit)

Change a specific element while preserving everything else.

Template:

"Using the provided image, change only the [specific element] to [new description]. Keep everything else in the image exactly the same, preserving the original style, lighting, and composition."

Examples:

• "Using the provided image, change only the blue car to a red vintage Mustang. Keep the street, buildings, and sky unchanged."
• "Remove the person standing in the background. Fill the area with the surrounding environment naturally."

Pattern 2: Style Transfer

Recreate the content in a different artistic style.

Template:

"Transform the provided [subject description] into the artistic style of [target style]. Preserve the original composition but render it with [description of stylistic elements]."

Examples:

• "Transform this city photograph into a Studio Ghibli anime scene with soft pastel colors and hand-painted textures."
• "Render this portrait as a Renaissance oil painting with dramatic chiaroscuro lighting and rich earth tones."

Pattern 3: Character in New Scene

Place a character from reference into a completely new environment.

Template:

"The same [character description from reference] from the reference image, now [action/pose] in [new environment]. Preserve the character's [key features to keep]. [Style and technical details]."

Examples:

• "The same silver-haired cyberpunk woman from the reference, now fighting in a neon-lit rain-soaked alley. Preserve her chrome dress, blue plasma lines, and cybernetic spine interface. Low-angle cinematic shot, 8k."

Pattern 4: Multi-Image Composition

Combine elements from multiple reference images.

Template:

"Create a new image by combining elements from the provided images. Take the [element from image 1] and [action] with the [element from image 2]. The final image should be [description of final scene]. Adjust lighting and shadows to create a cohesive result."

Examples:

• "Take the blue floral dress from the first image and put it on the woman from the second image. Generate a realistic full-body e-commerce fashion photo with studio lighting."
• "Place the product from image 1 into the kitchen scene from image 2, matching the lighting and perspective naturally."

Pattern 5: High-Fidelity Preservation

When critical details (face, logo, text) MUST be preserved exactly.

Template:

"Using the provided image(s), [edit description]. Ensure that [critical element] remains completely unchanged. The [modified element] should [integration description]."

Examples:

• "Add the logo from the second image onto her black t-shirt. Ensure the woman's face and features remain completely unchanged. The logo should look naturally printed on the fabric, following the folds."
• "Change the background to a tropical beach. The person's face, hair, clothing, and pose must remain pixel-perfect identical."

Pattern 6: Three-View / Character Sheet

Generate a character reference sheet from a single image.

Template:

"Create a professional character reference sheet showing [character from reference] in three views: front view, right side view, and back view. Neutral standing pose, pure white background, consistent proportions across all views, character design sheet style."

Tips

• Always write prompts in English for best quality, even when the user speaks another language.
• For edit_image, use the file path from canvas nodes or previous generations — never paste base64 data.
• When multiple images are needed from the same prompt, set n parameter on generate_image instead of calling multiple times.
• Multi-image editing: When the user references multiple images (e.g. "refer to the two character images I sent earlier"), use image_urls array. Up to 5 images supported.
• Single vs multi aspect ratio: Single-image edit always follows the input image's aspect ratio (cannot be overridden). Multi-image edit defaults to the first image but can be overridden with aspect_ratio.
• Iterative refinement: Use each edit output as the input for the next edit to progressively refine — describe only the incremental changes needed.
• Accurate text in images: To render text in generated images, put the desired text in quotes and describe its placement clearly (e.g. 'the word "HELLO" in bold serif font centered on the banner').