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 “帮我写提示词”, “优化提示词”, “给我一组出图更高级的 prompt”, “做成系列提示词”, “来点搞怪的”, “整点抽象图”, “做成会传播的图”
What this skill optimizes for
- Strong image type definition
- Clear visual hierarchy
- Specific layout and information modules
- Better conversion of abstract taste words into concrete visual directions
- Reusable prompt systems instead of one-off lines
- Strong contrast, absurdity, and meme potential when the goal is funny or viral content
- 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:
- Image type
- Core subject
- Composition / layout
- Supporting modules
- Visual tone
- Material / texture
- Typography / labeling
- Aspect ratio
Step 3: Translate vague taste words
Do not leave words like these unexplained:
- 高级感
- 电影感
- 氛围感
- 科技感
- 杂志感
Translate them into concrete visual instructions. Examples:
- 高级感 -> restrained layout, limited color palette, clean typography, premium materials, generous negative space
- 电影感 -> low-angle framing, dramatic lighting, foreground/background depth, emotional tension, reflective surfaces
- 科技感 -> glass panels, metal textures, interface modules, cool lighting, precision spacing
- 杂志感 -> 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: “要不要我直接用这条 Prompt 帮你生成一张图？”
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. 核心创意
Use 1-2 sentences to explain the visual concept in plain Chinese.
2. 完整 Prompt
Give one polished final prompt, usually as one continuous paragraph.
3. 为什么这样写
Explain briefly using 2-4 bullets. Focus on:
- 图像类型是否清楚
- 主体和版式是否明确
- 信息模块是否增强了层级
- 抽象审美词是否被翻译成了具体视觉语言
4. 可改写方向
Offer 2-3 short variation directions when useful, such as:
- 换时代 / 换城市
- 换职业 / 换情绪
- 换媒介形态（海报、UI、信息图、档案页）
- 换色彩和材质系统
5. 是否直接生图
如果当前环境配置了可用的生图工具或生图模型，例如 Hermes 的 image_generate 工具，在完成 Prompt 输出后，用一句话询问用户是否要直接生成图片： “要不要我直接用这条 Prompt 帮你生成一张图？”
只有在用户明确要求“直接生成/帮我出图/生成一张”或用户回答确认后，才调用生图工具。不要在用户只要求写 Prompt 时自动生图。
如果环境没有可用的生图能力，不要追加这个确认问题，也不要暗示可以在当前环境直接生成。
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:
分类标题
Name the category first, especially when the batch is large. Examples:
- 历史世界 × 现代界面
- 信息图 × 情绪表达
- 品牌提案 × 商业视觉
- 世界观档案页
- 城市观察 × 社会情绪
每条 Prompt 的默认格式
For each prompt, use:
[编号]. [标题]
- 核心创意：用一句话讲清楚这个画面想做什么
- 完整 Prompt：给一条可直接使用的完整 Prompt
- 可替换变量：列 2-4 个最值得替换的变量，方便用户自己扩写
Batch-size guidance
- 1-5 条：可以保留“为什么这样写”
- 6-20 条：以“核心创意 + 完整 Prompt + 可替换变量”为主，保持紧凑
- 20 条以上：优先按分类分组，减少逐条解释，除非用户明确要求详细拆解
Default consistency rules
- 同一批内容尽量保持统一结构
- 同一分类优先共享一套视觉逻辑，再替换主题变量
- 如果是做资料包，优先保证可复制、可扩写、可分类整理
- 如果是做灵感清单，可以适当放宽结构，但仍然保留图像类型和版式意识
Upgrade path for weak prompts
If the user gives a weak prompt, rewrite it by upgrading in this order:
- clarify image type
- define subject
- define composition
- add modules
- translate vague taste words
- add materials / labels / typography
- choose aspect ratio
Chinese example response
核心创意
做一张“城市失眠指数”主题的信息图海报，把深夜情绪转成可以被观看的数据地图。
完整 Prompt
“城市失眠指数”信息图海报，中心是一张俯视夜景地图，按照 22:00、00:00、02:00、04:00 四个时段分层发光，商业区高亮，住宅区以昏黄窗光点阵呈现，四周嵌入咖啡销量、夜间打车热度、音乐播放峰值、社交媒体活跃度四个圆形数据模块，标题像国际杂志专题页，副标题写“谁还醒着，谁在假装睡着”，整体冷静克制，带轻微都市焦虑感，色彩控制在霓虹紫、电蓝、夜黑和少量暖黄，比例 4:5。
为什么这样写
- 先定义为“信息图海报”，模型会按专题视觉去组织画面
- 中心地图 + 四周模块，画面层级会更清楚
- “都市焦虑感”被拆成夜景、冷色、高亮区域这些具体视觉指令
- 配色和比例明确后，更容易出成品感
可改写方向
- 换成“加班热力图”主题
- 换成“地铁情绪地图”结构
- 换成杂志封面而不是信息图海报
Chinese batch example
历史世界 × 现代界面
1. 唐朝人的外卖 App
- 核心创意：把唐朝日常生活翻译成一张现代外卖首页界面，让历史感和产品感同时成立。
- 完整 Prompt：唐朝人的外卖 App，画面模拟手机外卖首页界面，顶部定位显示“长安·平康坊”，推荐位展示胡饼、炙羊肉、葡萄酿，商家头像采用工笔画掌柜半身像，评分以铜钱图标呈现，底部导航栏完整保留现代产品结构，状态栏显示“大唐信号满格”和“开元二十四年”，整体配色为赭石、石绿、金箔红，字体融合碑刻感标题字与细无衬线，界面像真实产品设计稿，同时带有历史穿越幽默感，比例 9:16。
- 可替换变量：朝代、城市、品类、界面类型
2. 明代人的求职网站首页
- 核心创意：把古代职业体系做成现代招聘平台首页，重点放在职位卡片和筛选逻辑。
- 完整 Prompt：明代人的求职网站首页，历史服饰与现代招聘平台界面融合，主界面有职位推荐“修史官、书院讲郎、船务账房、织造监督、宫廷画师”，筛选条件写“擅诗文、通算学、善骑射、可外派”，人物简历卡使用工笔肖像，薪资单位以俸禄石米表示，整体排版像真实招聘网站首页，比例 16:9。
- 可替换变量：时代、岗位、筛选条件、薪资表达
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 “高级感” 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.