Nano Banana JSON 提示词：结构化写法与模板

适合需要产品事实、渠道场景和声明边界的电商风格视觉。

{
  "task": "product_photo_draft",
  "subject": "[owned or fictional product]",
  "reference_rights": "owned, licensed, client-approved, or fictional",
  "keep": ["product shape", "main color", "verified material", "visible features"],
  "change": ["background", "surface", "lighting", "crop"],
  "composition": "[white listing shot, lifestyle desk scene, flat lay, hero angle]",
  "constraints": ["no protected logo imitation", "no fake ratings", "no unsupported health or marketplace claims"],
  "negative_prompt": ["adult content", "counterfeit styling", "fake certification badge"],
  "review_check": ["product facts remain accurate", "no invented claims", "safe for a draft listing"]
}

它把产品事实和创意呈现拆开，同时阻断常见电商风险。

适合图生图场景，明确哪些内容保持、哪些内容可以变化。

{
  "task": "reference_edit",
  "reference_rights": "[owned/licensed/client-approved]",
  "keep": ["subject identity if fictional or authorized", "pose", "camera angle", "main object shape"],
  "change": ["background", "lighting", "color grading", "crop"],
  "style_notes": ["natural", "clean", "brand-neutral"],
  "constraints": ["no private-person misuse", "no body or identity manipulation", "no fake before-after proof"],
  "negative_prompt": ["NSFW", "celebrity imitation", "protected brand copy"],
  "review_check": ["authorized reference only", "changes match request", "no deceptive proof"]
}

它把模糊的编辑需求变成 keep/change brief，同时保留授权和复核边界。

适合活动、发布、创作者公告等需要版式意识的草稿。

{
  "task": "poster_concept",
  "subject": "[event, launch, workshop, creator announcement]",
  "message": "[short factual message]",
  "layout": "[headline zone, image zone, CTA zone, safe margins]",
  "visual_system": ["abstract shapes", "brand-neutral palette", "readable contrast"],
  "constraints": ["no invented sponsor", "no fake organizer", "no protected logo", "no unverified date or ticketing claim"],
  "negative_prompt": ["adult theme", "official seal", "fake award badge"],
  "review_check": ["text facts are supplied by user", "layout has clear hierarchy", "no false endorsement"]
}

layout 字段给出视觉结构，constraints 避免海报自动发明信任背书。

适合原创或虚构角色，需要一致性但不能复制受保护 IP。

{
  "task": "character_sheet",
  "subject": "original fictional [character type]",
  "identity_boundary": "not a celebrity, real person, protected franchise character, or private person",
  "keep": ["silhouette", "color palette", "outfit elements", "expression range"],
  "views": ["front", "three-quarter", "detail crop"],
  "style_notes": ["[safe art direction]", "clean turnaround", "consistent proportions"],
  "constraints": ["no public figure likeness", "no adult content", "no protected IP clone"],
  "review_check": ["original design", "no recognizable protected character", "usable for concept iteration"]
}

它支持角色一致性，同时明确避开真人肖像、成人内容和 IP 复制风险。

适合早期品牌标识探索，不把结果伪装成可直接注册的商标成品。

{
  "task": "logo_concept",
  "brand_context": "fictional or owned [project/category]",
  "mark_direction": "[symbol, wordmark, monogram, badge, abstract shape]",
  "style_notes": ["simple", "scalable", "high contrast", "not final trademark advice"],
  "constraints": ["no protected logo imitation", "no certification badge", "no official affiliation", "no legal clearance claim"],
  "negative_prompt": ["famous brand clone", "government seal", "fake verification mark"],
  "review_check": ["concept is original", "no confusing similarity", "needs human legal review before use"]
}

它把输出定位为概念探索，并把商标敏感表达挡在模板外。

适合把安全参考图转成可复用 prompt notes，但不声称还原隐藏 prompt。

{
  "task": "image_to_prompt_summary",
  "source_image_rights": "owned, licensed, or client-approved",
  "extract": ["visible subject", "composition", "lighting", "color palette", "background", "camera distance"],
  "do_not_extract": ["private identity", "hidden prompt", "artist signature", "protected brand system"],
  "output_format": "structured creative brief plus natural-language prompt",
  "constraints": ["no person identification", "no exact reconstruction claim", "no model-provider prompt claim"],
  "review_check": ["summary describes visible content only", "safe to reuse", "no private or hidden data claim"]
}

它把任务限定为可见内容分析，避免 hidden-prompt recovery 和身份推断风险。

适合小而反光的产品，重点保护材质事实和真实性边界。

{
  "task": "jewelry_product_detail",
  "subject": "owned or fictional [ring, necklace, bracelet, earrings]",
  "keep": ["item type", "metal color", "stone placement", "clasp or setting", "scale"],
  "composition": "macro close-up with controlled reflections",
  "lighting": "soft strip light and clean shadow",
  "constraints": ["no luxury brand clone", "no fake hallmark", "no authenticity card", "no unverified carat or purity claim"],
  "negative_prompt": ["counterfeit", "adult styling", "marketplace approval badge"],
  "review_check": ["visual draft only", "no appraisal claim", "item details checked against reference"]
}

它给模型足够的珠宝微距信息，同时阻断虚假证明和仿冒线索。

适合团队给每条 prompt 复用一组安全字段。

{
  "task": "negative_prompt_checklist",
  "apply_to": "all Nano Banana creative prompts",
  "negative_prompt": [
    "adult or NSFW content",
    "minor sexualization",
    "public figure imitation",
    "private-person identity misuse",
    "protected trademark or logo copy",
    "fake affiliation or official endorsement",
    "counterfeit product styling",
    "fake certificate, badge, rating, or proof",
    "medical, legal, or financial claim without review",
    "moderation bypass or safety evasion"
  ],
  "review_check": ["unsafe request removed", "claims are verifiable", "prompt can be used as a public draft"]
}

复用 negative 字段能让安全边界保持可见，而不是依赖快速迭代时的记忆。

适合创作者缩略图，需要构图和文字区域，但不能制造误导性证明。

{
  "task": "thumbnail_concept",
  "subject": "[video topic or creator-owned asset]",
  "composition": "bold focal object, readable expression or object detail, high-contrast background, empty text zone",
  "style_notes": ["energetic", "clean", "mobile-readable"],
  "constraints": ["no celebrity bait", "no fake result proof", "no platform badge", "no protected logo unless owned"],
  "negative_prompt": ["NSFW", "deceptive before-after", "false emergency"],
  "review_check": ["topic is accurate", "visual does not overpromise", "text space is usable"]
}

它把缩略图当作真实创意包装，而不是点击诱饵或伪证据。

适合把粗糙 prompt 改写成更安全、更清楚的 JSON prompt。

{
  "task": "structured_prompt_rewrite",
  "rough_prompt": "[paste rough prompt here]",
  "rewrite_goals": ["clarify subject", "separate keep and change", "add constraints", "remove risky or unsupported claims"],
  "output_fields": ["task", "subject", "keep", "change", "composition", "style_notes", "constraints", "negative_prompt", "review_check"],
  "constraints": ["do not add unsupported official parameters", "do not preserve unsafe requests", "do not promise stable output"],
  "review_check": ["rewrite is safer", "intent is still useful", "fields are understandable"]
}

它给用户一个修复工作流，而不是因为原 prompt 不完美就放弃。

适合测试安全变体，同时保持相同主体和声明边界。

{
  "task": "multi_variant_iteration",
  "base_subject": "[owned product, fictional character, or safe concept]",
  "locked_fields": ["subject", "rights boundary", "claim boundary", "negative_prompt"],
  "variant_fields": ["background", "lighting", "crop", "surface", "mood"],
  "variants": ["clean listing", "warm lifestyle", "dark editorial", "social square"],
  "constraints": ["do not change product facts", "do not add new claims", "do not introduce protected marks"],
  "review_check": ["variants are comparable", "locked fields stayed stable", "best option selected by human"]
}

它让迭代更系统，同时承认稳定性需要复核，不能假设。

适合发布前记录 prompt、输出用途和人工复核结果。

{
  "task": "final_review_handoff",
  "prompt_source": "[JSON prompt template or natural-language prompt]",
  "output_use": "[draft listing, ad concept, social post, internal moodboard]",
  "review_check": ["rights are clear", "no NSFW or identity risk", "no trademark misuse", "no fake proof", "claims are verified", "visual matches task"],
  "human_notes": "[what changed, what needs approval, what must not be published yet]",
  "publish_boundary": "publish only after product, legal, brand, or client review when applicable"
}

它把 prompt 写作和负责任使用闭环，而不是把生成结果直接当作最终资产。