giggle-official/skills/giggle-gpt-image-2

GPT-Image-2 Prompt & Generation Assistant

Helps users generate high-quality prompts for GPT-Image-2, then directly calls the Giggle API using gpt-image-2-fast to generate images.

Workflow

Step 1: Understand the Request

Quickly extract the following key information (skip what is already provided; ask only 1–2 follow-up questions for the most critical missing pieces):

• Subject / Content: What to draw? What are the core elements?
• Use Case: Poster / portrait / UI / social media cover / character design, etc. (see categories below)
• Style Preference: Realistic / illustration / cyberpunk / Chinese ink / cinematic, etc.
• Aspect Ratio: 9:16 portrait / 16:9 landscape / 1:1 square (default: 9:16)
• Image-to-Image: Is there a reference image? (If yes, specify which aspects to preserve)

If the user has already provided enough information, proceed directly to the next steps without asking again.

Step 2: Match Scene Type and Load Reference Examples

Match the user's request to one of the categories below, then read the corresponding file for inspiration from original cases:

| Scene Type | Reference File | Key Styles | |-----------|---------------|-----------| | Portrait / Photography | references/examples-portrait.md | Film look, Korean idol, influencer shots, CCD aesthetic (18 cases) | | Poster / Illustration | references/examples-poster.md | City promo, science infographic, dark epic, watercolor illustration (46 cases) | | Character Design | references/examples-character.md | Game characters, anime character sheets, emoji/sticker series (9 cases) | | UI / Social Media Mockup | references/examples-ui-social.md | TikTok screenshots, UI design systems, social feeds (25 cases) | | Creative Mix | references/examples-community.md | Style transfer, game screenshots, historical crossover (15 cases) |

Universal Prompt Templates (slot-based structural frameworks): references/prompt-templates.md

Step 3: Build the Prompt

Assemble using the following structure (adjust emphasis based on scene type):

Portrait / Photography:

[Capture medium + style] [Lighting atmosphere] [Subject description: age/appearance/outfit] [Pose/expression] [Background scene] [Technical specs] [Negative prompts]

Poster / Illustration:

[Style definition] [Composition: S-curve / diagonal / symmetry] [Hero visual subject] [Background scene + landmarks] [Color palette] [Typography / text] [Aspect ratio] [Quality modifiers]

Character Design:

[World-setting + character role] [Appearance + outfit + equipment] [Multi-view notes] [Color reference] [Detail requirements] [Background / layout]

UI / Mockup:

[Platform name + interface type] [Content details: text / username / content] [UI element description] [Character in scene] [Aspect ratio]

Step 4: Output the Final Prompt First

Finalize the prompt before calling the Giggle API.

Output format:

1. Final Prompt (ready to copy) Wrap in a code block:

[complete prompt]

2. Brief Notes (3 lines max)

• Structure / style used
• Key optimizations
• Whether this will be text-to-image or image-to-image

Step 5: Call the Giggle API to Generate the Image

After generating the prompt, continue to image generation — do not stop at the prompt stage.

5.1 Mode Selection

• No reference image: Call text-to-image endpoint POST /api/v1/generation/text-to-image
• Has reference image: Call image-to-image endpoint POST /api/v1/generation/image-to-image
• Model is fixed: gpt-image-2-fast

5.2 Reference Image Input Rules

• Prefer user-provided local file paths
• Also supports remote image URLs
• For local paths, the script scripts/generate_gpt_image.py handles base64 conversion automatically
• For URLs, pass directly as reference_images[].url

5.3 Run the Script

Use the bundled script:

python scripts/generate_gpt_image.py \
  --prompt "<final prompt>" \
  --aspect-ratio <ratio> \
  --output-format kv

With a reference image:

python scripts/generate_gpt_image.py \
  --prompt "<final prompt>" \
  --aspect-ratio <ratio> \
  --reference-image "<local path or remote URL>" \
  --output-format kv

Optional arguments:

• --count: Number of images to generate, default 1
• --timeout: Maximum wait time in seconds, default 300
• --output-format: kv / json / plain, default kv

5.4 Response to User

The script automatically:

• Reads GIGGLE_API_KEY
• Submits the task
• Polls /api/v1/generation/task/query
• Extracts result image URLs from data.urls per the query API spec
• Falls back to other nested URL fields for backwards compatibility
• Outputs results using fixed key names to reduce the chance of LLMs missing links:

RESULT_STATUS=success
RESULT_PRIMARY_URL=https://...
RESULT_URL_COUNT=1
RESULT_URL_1=https://...

When replying to the user:

• Briefly state whether text-to-image or image-to-image was used
• Provide the final prompt
• Use RESULT_PRIMARY_URL as the primary result
• Return the image URL as-is

Important requirements:

• Do not wrap the URL in a markdown link
• Do not truncate the URL
• Do not remove signature parameters
• Do not rewrite success results as "click here to view"

Core Techniques

Quality Boosters (add as needed):

• Photographic feel: 35mm film, film grain, cinematic, photorealistic, 8K
• Illustration feel: high detail, masterpiece, professional illustration
• Composition control: --ar 9:16, 9:16 vertical, 16:9 horizontal

Language Tips:

• Chinese prompts work better for Chinese aesthetic styles and Chinese typography
• English prompts work better for Western portraits and cinematic quality
• When mixing: use Chinese for the subject description, English for technical terms

Image-to-Image (Reference Images):

• Style reference: in the style of [reference image], maintain the color palette
• Subject reference: same person, consistent facial features
• Partial edits: describe what to preserve first, then specify what to change
• If the user says "keep the subject, change only the background/outfit/material/lighting," put the "preserve" items in the first half of the prompt

Negative Prompts (add as appropriate):

no watermark, no text, no plastic skin, no over-sharpening, no blur, no deformed hands

Runtime Environment

Requires system environment variable:

export GIGGLE_API_KEY=your_api_key

Giggle API Key is available at https://giggle.pro/developer.

Reference Resources

Detailed prompt examples are organized in references/examples-portrait.md, references/examples-poster.md, references/examples-character.md, references/examples-ui-social.md, and references/examples-community.md. Consult the relevant category file when you need inspiration.