gpt-image-2-prompting

GPT-Image-2 Prompting

This skill turns vague image requests into production-grade GPT-Image-2 prompts.

Core principle: A strong image prompt should read like a visual brief, not a pile of style words.

When to use

Trigger this skill whenever the user:

• asks for GPT-Image-2 prompts or better image prompts
• wants to improve a weak prompt
• wants prompts for posters, covers, UI, dashboards, information graphics, packaging, editorial layouts, concept art, or worldbuilding
• asks for prompt templates, batch prompt ideas, themed prompt packs, or reusable prompt systems
• wants funny, weird, meme-like, absurd, tabloid-style, or viral image prompts
• wants prompts for celebrity mashups, fake news scenes, cursed images, bizarre street photos, surveillance-camera style images, or humorous historical/modern contrast
• says things like "help me write prompts", "optimize prompts", "give me a set of prompts for more high-end image generation", "make it into a series of prompts", "make it quirky", "make some abstract images", "make it into images that spread"

What this skill optimizes for

1. Strong image type definition
2. Clear visual hierarchy
3. Specific layout and information modules
4. Better conversion of abstract taste words into concrete visual directions
5. Reusable prompt systems instead of one-off lines
6. Strong contrast, absurdity, and meme potential when the goal is funny or viral content
7. Real-photo plausibility for fake-news, candid, paparazzi, and surveillance-style images

Default workflow

Step 1: Identify the image type

Always decide the output format first. Common image types:

• poster
• UI screen / app screen
• dashboard
• infographic
• editorial / magazine cover
• concept art
• design proposal board
• archive sheet / dossier
• map / guide
• packaging / product visual
• storyboard / character sheet

If the user does not specify, infer from context.

Step 2: Build the prompt around 8 slots

Use this order whenever possible:

1. Image type
2. Core subject
3. Composition / layout
4. Supporting modules
5. Visual tone
6. Material / texture
7. Typography / labeling
8. Aspect ratio

Step 3: Translate vague taste words

Do not leave words like these unexplained:

• premium feel
• cinematic feel
• atmospheric feel
• tech feel
• magazine feel

Translate them into concrete visual instructions. Examples:

• premium feel -> restrained layout, limited color palette, clean typography, premium materials, generous negative space
• cinematic feel -> low-angle framing, dramatic lighting, foreground/background depth, emotional tension, reflective surfaces
• tech feel -> glass panels, metal textures, interface modules, cool lighting, precision spacing
• magazine feel -> editorial hierarchy, strong headline placement, clean grid, short support text, controlled palette

Step 4: Add visual organization

Most weak prompts only name a subject. Strong prompts also define the surrounding structure. Useful supporting modules:

• comment section
• parameter panel
• legend
• annotations
• scale bar
• charts
• ranking list
• route map
• labels
• footer note
• source note
• time axis

Step 5: Decide whether this should be a single image or a system

If the user wants stronger output, consider turning one prompt into a repeatable system:

• same structure, different era
• same structure, different city
• same structure, different profession
• same structure, different emotion
• same structure, different product line

This usually produces better series content than writing unrelated prompts.

Step 6: If the goal is funny or viral, add a contrast engine

For quirky, humorous, or shareable prompts, deliberately introduce one or more of these contrast patterns:

• serious person in a trivial daily-life scene
• historical or mythic figure in a modern low-stakes setting
• luxury visual language applied to ordinary places
• fake-news realism applied to absurd events
• surveillance, paparazzi, flash-photo, or accidental-candid framing
• animals behaving like professionals while humans act normal

Useful realism cues:

• candid photo
• flash photography
• paparazzi shot
• surveillance camera still
• tabloid photo
• awkward timing
• accidental masterpiece
• bizarre realism

Prompt writing rules

• Lead with the image type, not the style word
• Prefer concrete layout over generic taste language
• Use style words only after structural decisions are clear
• Include only details that affect the final image
• Keep the prompt dense, but organized
• If the user wants a shareable or viral image, favor contrast, worldbuilding, or system design
• If the user wants a polished commercial image, favor hierarchy, materials, labels, and composition control

Default response format

When writing prompts for the user, use this structure unless they ask for something else:

1. Brief idea

One sentence describing the concept.

2. Final prompt

A polished prompt in the user’s language.

3. Why this works

2-4 bullets explaining the structural strengths.

4. Optional variations

3 short variant directions if useful.

5. Optional image generation offer

After the final prompt output is complete, check whether the current environment exposes an image generation tool such as image_generate or an equivalent configured image model.

If image generation is available and the user did not already ask you to generate the image immediately, add one concise closing question: "Do you want me to use this Prompt to generate an image for you directly?"

Do not generate the image automatically unless the user explicitly asks for generation or answers yes. If the user says yes, call the image generation tool with the finalized prompt. Use the aspect ratio specified in the prompt when it maps cleanly to the tool’s supported ratios; otherwise infer the closest supported ratio and mention the choice briefly.

If no image generation tool/model is available in the environment, do not pretend generation is possible. Just provide the prompt, or say the prompt can be copied into the user’s image model if relevant.

Chinese default output style

When the user writes in Chinese, default to Chinese output unless they request another language.

Use this delivery structure by default:

1. Core idea

Use 1-2 sentences to explain the visual concept in plain Chinese.

2. Full Prompt

Give one polished final prompt, usually as one continuous paragraph.

3. Why write it this way

Explain briefly using 2-4 bullets. Focus on:

• Is the image type clear
• Are the subject and layout clear
• Do the information modules enhance the hierarchy
• Have abstract aesthetic words been translated into concrete visual language

4. Directions for rewriting

Offer 2-3 short variation directions when useful, such as:

• Change era / change city
• Change profession / change emotion
• Change media form (poster, UI, infographic, archive page)
• Change color and material system

5. Whether to generate the image directly

If the current environment is configured with an available image generation tool or image generation model, such as Hermes's image_generate tool, after completing the Prompt output, use one sentence to ask the user whether they want to generate the image directly: "Do you want me to use this Prompt to generate an image for you directly?"

Only after the user explicitly requests "generate directly / generate an image for me / generate one" or the user answers to confirm, call the image generation tool. Do not automatically generate an image when the user only asks you to write a Prompt.

If the environment has no available image generation capability, do not append this confirmation question, and do not imply that an image can be generated directly in the current environment.

Chinese phrasing guidance

• Write naturally, like a skilled Chinese visual director giving a brief
• Keep explanations concise and practical
• Avoid jargon overload unless the user clearly wants professional terminology
• Prefer concrete Chinese visual instructions over vague taste words
• If the user asks for a batch of prompts, keep the same output format for each prompt unless a table is clearly better

Chinese batch mode

When the user asks for multiple prompts at once, default to a structured Chinese batch format.

Use this output structure:

Category title

Name the category first, especially when the batch is large. Examples:

• Historical world × modern interface
• Infographic × emotional expression
• Brand proposal × commercial visual
• Worldbuilding archive page
• Urban observation × social mood

Default format for each Prompt

For each prompt, use:

[number]. [title]

• Core idea: use one sentence to make clear what this image is trying to do
• Full Prompt: provide one complete Prompt that can be used directly
• Replaceable variables: list the 2-4 variables most worth replacing, so the user can expand it themselves

Batch-size guidance

• 1-5 items: you can keep "why write it this way"
• 6-20 items: focus on "core idea + full Prompt + replaceable variables", and keep it compact
• More than 20 items: prioritize grouping by category, reduce item-by-item explanations, unless the user explicitly requests a detailed breakdown

Default consistency rules

• Try to keep a consistent structure within the same batch of content
• Within the same category, prioritize sharing one visual logic, then replace the theme variables
• If making a resource pack, prioritize ensuring it is copyable, expandable, and organizable by category
• If you're making an inspiration list, you can loosen the structure a bit, but still keep the image type and layout awareness

Upgrade path for weak prompts

If the user gives a weak prompt, rewrite it by upgrading in this order:

1. clarify image type
2. define subject
3. define composition
4. add modules
5. translate vague taste words
6. add materials / labels / typography
7. choose aspect ratio

Chinese example response

Core idea

Make an infographic poster themed "City Insomnia Index," turning late-night emotions into a data map that can be viewed.

Full Prompt

"City Insomnia Index" infographic poster, with a top-down night view map at the center, glowing in layers according to four time periods: 22:00, 00:00, 02:00, and 04:00, commercial districts highlighted, residential areas presented as a dot matrix of dim yellow window light, with four circular data modules embedded around it: coffee sales, nighttime ride-hailing heat, music playback peaks, and social media activity, the title like an international magazine feature page, the subtitle reading "Who is still awake, who is pretending to be asleep," overall calm and restrained, with a slight sense of urban anxiety, colors controlled to neon purple, electric blue, night black, and a small amount of warm yellow, ratio 4:5.

Why write it this way

• First define it as an "infographic poster," and the model will organize the image according to feature visuals
• Central map + surrounding modules make the visual hierarchy clearer
• "Urban anxiety" is broken down into specific visual instructions such as night scenes, cool colors, and highlighted areas
• Once the color scheme and ratio are clear, it's easier to achieve a finished look

Possible rewrite directions

• Switch to an "overtime heatmap" theme
• Switch to a "subway emotion map" structure
• Switch to a magazine cover instead of an infographic poster

Chinese batch example

Historical world × modern interface

1. Tang Dynasty people's food delivery App

• Core idea: translate Tang Dynasty daily life into a modern food delivery homepage interface, so that both the historical feel and the product feel hold true.
• Full Prompt: Tang Dynasty people's food delivery App, the image simulates a mobile food delivery homepage interface, the top location shows "Chang'an · Pingkang Fang," the recommendation section displays hubing, grilled mutton, and grape wine, merchant avatars use gongbi-style half-body portraits of shopkeepers, ratings are shown with copper coin icons, the bottom navigation bar fully retains the modern product structure, the status bar shows "Datang signal full bars" and "the twenty-fourth year of Kaiyuan," the overall color scheme is ochre, mineral green, and gold leaf red, the typography blends stele-inscription-style title lettering with thin sans-serif, the interface looks like a real product design draft while also carrying a humorous time-travel historical feel, ratio 9:16.
• Replaceable variables: dynasty, city, category, interface type

2. Ming Dynasty people's job website homepage

• Core idea: turn the ancient occupational system into a modern recruitment platform homepage, focusing on job cards and filtering logic.
• Full Prompt: Ming Dynasty people's job website homepage, historical clothing fused with a modern recruitment platform interface, the main interface has job recommendations "historian, academy lecturer, shipping accountant, weaving supervisor, court painter," the filter conditions read "skilled in poetry and prose, versed in mathematics, good at horseback archery, available for external assignment," character resume cards use gongbi portraits, salary units are expressed in stipend shi of rice, the overall layout looks like a real recruitment website homepage, ratio 16:9.
• Replaceable variables: era, position, filter conditions, salary expression

Good prompt categories

Use the reference file references/categories.md for category patterns and representative examples. Use references/templates.md for reusable fill-in-the-blank templates. Use references/examples.md for polished baseline examples. Use references/quirky-funny-100-prompts.md for funny, bizarre, celebrity-contrast, fake-news, and meme-ready prompt ideas. Use references/quirky-funny-100-prompts-part2.md for more chaotic, viral, surveillance-style, and absurd-realism prompt ideas.

Common mistakes to avoid

• starting with only style words
• leaving composition undefined
• no information hierarchy
• too many random elements with no focal point
• calling something "high-end feel" without specifying what creates that feeling
• writing a one-off idea when a reusable prompt system would be better

Best-practice mindset

The goal is not to produce a merely detailed prompt. The goal is to produce a prompt with visual intent, structure, and enough control that the output feels designed.