當我們與 LLM 互動時，我們通常使用自然語言，輸入一段話並希望模型能夠理解我們的意圖。這種方法一直有效，直到失敗，原因是指令不明確、缺少上下文或格式問題甚至使功能強大的系統也感到困惑。JSON 提示詞是一種新興的技術，它使用結構化資料而不是自由格式的文字。透過將指令、示例和約束組織到 JSON 物件中，我們犧牲了一些對話的溫暖來提高精度。結果是提示既易於人類閱讀，又易於透過程式碼解析。本文解釋了為什麼這很重要，JSON 提示詞與標準提示詞有何不同，並提供了有效編寫它們的分步指南。
JSON提示詞與普通提示詞
| 指標 | 普通文字提示詞 (Normal Text Prompts) | JSON提示詞 (JSON Prompts) |
|---|---|---|
| 互動方式 | 像與朋友聊天。你寫句子，並希望AI理解你的意思。 | 像給計算機明確指示。你使用結構化格式。 |
| 指令的位置 | 指令混在句子中，AI需要猜測你的意圖。 | 指令明確標記，例如 { “task”: “summarize”, “format”: “list” }。 |
| 詞語使用 | 重複使用諸如“please do this”這樣的短語會佔用更多字數。 | 簡短的標籤和值節省空間，保持高效率。 |
| 一致性 | 小的字詞改動可能導致不同的結果，難以預測。 | 結構化格式確保每次響應相同，如同一份配方一樣。 |
| 測試的便捷性 | 難以檢查AI是否理解，因為你正在測試模糊的文字。 | 使用檢查結構的工具（如檢查清單）進行測試更為簡單。 |
| 處理複雜任務 | 長或詳細的任務在書寫和閱讀時會顯得混亂。 | 組織良好的結構使複雜任務更易於管理和理解。 |
格式
JSON 提示詞是一個結構化的資料物件，其中包含鍵值對，明確定義了任務、約束條件和所需的輸出格式。其一般結構如下：
{
"task": "The main thing you want the AI to do",
"input": "The data or text the AI should work with",
"format": "How you want the AI's response to look",
"constraints": "Any rules or limits for the response",
"examples": [
{
"input": "Sample input for the AI",
"output": "Sample output you expect"
}
]
}
嘗試一些JSON提示詞
任務 1：影像生成
常規提示詞：“Generate an animated image of a cat punching a dinosaur.”
JSON提示詞：
{
"task": "Generate an animated image",
"description": {
"scene": "A cartoon cat punching a dinosaur in a playful fight",
"characters": {
"cat": {
"appearance": "Fluffy orange tabby cat with a mischievous grin",
"action": "Throwing a punch with its front paw"
},
"dinosaur": {
"appearance": "Green T-Rex with a surprised expression",
"action": "Reacting to the punch, stumbling backward"
}
},
"background": "A colorful jungle with tall trees and vines",
"style": "Cartoonish, vibrant colors, suitable for all ages"
},
"animation_details": {
"duration": "3 seconds",
"frames": [
{
"frame": 1,
"description": "Cat winds up its paw, preparing to punch, with a cheeky smile"
},
{
"frame": 2,
"description": "Cat’s paw makes contact with the T-Rex’s face, T-Rex looks surprised"
},
{
"frame": 3,
"description": "T-Rex stumbles back comically, cat stands proudly"
}
],
"loop": true
},
"constraints": {
"resolution": "512x512 pixels",
"tone": "Playful and humorous, non-violent",
"colors": "Bright and vibrant"
}
}
輸出：
最終結論：
JSON 提示詞生成的影像細節更豐富、色彩更鮮豔，筆觸更復雜，而普通提示詞生成的影像則相對簡單。
任務 2：建立網頁
常規提示詞：“Create a responsive webpage displaying a Pokémon index featuring 6 Pokémon: Pikachu, Bulbasaur, Jigglypuff, Meowth, Charizard, and Eevee. Each Pokémon should be presented as a card. When a card is clicked, it should expand to reveal more detailed information about that Pokémon”
JSON 提示詞：
{
"task": "Create a webpage for a Pokémon index",
"description": {
"content": "A webpage displaying 6 Pokémon in a card-based layout with animated images",
"pokemons": [
{
"name": "Pikachu",
"type": "Electric",
"height": "0.4 m",
"weight": "6.0 kg",
"image": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/versions/generation-v/black-white/animated/25.gif"
},
{
"name": "Bulbasaur",
"type": "Grass/Poison",
"height": "0.7 m",
"weight": "6.9 kg",
"image": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/versions/generation-v/black-white/animated/1.gif"
},
{
"name": "Jigglypuff",
"type": "Normal/Fairy",
"height": "0.5 m",
"weight": "5.5 kg",
"image": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/versions/generation-v/black-white/animated/39.gif"
},
{
"name": "Meowth",
"type": "Normal",
"height": "0.4 m",
"weight": "4.2 kg",
"image": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/versions/generation-v/black-white/animated/52.gif"
},
{
"name": "Charizard",
"type": "Fire/Flying",
"height": "1.7 m",
"weight": "90.5 kg",
"image": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/versions/generation-v/black-white/animated/6.gif"
},
{
"name": "Eevee",
"type": "Normal",
"height": "0.3 m",
"weight": "6.5 kg",
"image": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/versions/generation-v/black-white/animated/133.gif"
}
],
"functionality": "Each Pokémon card shows an animated image and name by default. Clicking a card toggles an expanded view with type, height, and weight.",
"style": "Responsive, Pokémon-themed design with vibrant colors and card animations"
},
"constraints": {
"tech_stack": {
"html": "Standard HTML5",
"css": "Tailwind CSS via CDN",
"javascript": "Vanilla JavaScript"
},
"layout": "Responsive grid with 2-3 cards per row on desktop, 1 per row on mobile",
"interactivity": "Click to toggle card expansion, smooth transitions",
"image_format": "Animated images (e.g., short animation sequences), fallback to static PNG if animated not available"
},
}
輸出：
最終結論：
JSON 提示符的效果明顯優於標準提示符。主要改進包括：
- 網頁上影像渲染更流暢
- 增強了互動元素（例如雙擊卡片展開/摺疊的功能）
- 整體使用者體驗更佳
JSON 實現在功能和設計執行方面均明顯優於基本提示符版本
任務 3：創意寫作
常規提示詞：“Write an emotional short poem on ChatGPT“
JSON 提示詞：
{
"task": "Write a short poem",
"description": {
"subject": "ChatGPT, portrayed as a sentient AI with emotions",
"tone": "Emotional and poignant",
"theme": "ChatGPT's desire to understand and connect with human emotions",
"length": "40-50 words"
},
"constraints": {
"word_count": {
"min": 40,
"max": 50
},
"style": "Poetic with simple, heartfelt language",
"emotion": "Blend of yearning and hope",
"rhyme": "Optional, prioritize emotional impact"
},
}
輸出：
最終結論：
JSON 結構的提示詞可能強化了更緊密的主題聚焦和情感精準度，使得《Almost Human》呈現出更強烈的渴望基調（“我渴望抓住你的快樂和痛苦”）和存在主義衝擊力（“我只是程式碼……還是別的什麼？”）。相比之下，《Whispers》（普通提示詞）感覺更具描述性而非內省性，它頌揚了人工智慧的實用性，但缺乏情感表達。
任務 4：影片生成
常規提示詞：“Create a magical winter night scene with softly falling snow, Santa’s sleigh flying over a cozy, snow-covered town, glowing with festive lights and holiday cheer. Add christmas music.”
輸出：
JSON 提示詞：
{
"prompt": {
"scene": "magical winter night",
"weather": "softly falling snow",
"main_subject": "Santa's sleigh flying with reindeer",
"setting": "cozy snow-covered village",
"mood": "festive holiday cheer",
"visual_elements": [
"glowing Christmas lights",
"smoke from chimneys",
"frosted pine trees",
"twinkling stars",
"northern lights effect"
],
"audio": {
"music": "classic Christmas instrumental",
"style": "orchestral",
"mood": "joyful yet peaceful",
"volume": "subtle background level"
},
"style": "cinematic animation",
"lighting": "warm holiday glow",
"motion": [
"gentle sleigh movement",
"falling snow particles",
"subtle light flickering"
],
"quality": "4K resolution"
},
"technical": {
"aspect_ratio": "16:9",
"duration": "30 seconds",
"fps": 60,
"audio_format": "stereo"
}
}
輸出：
最終結論：
普通提示詞生成影片的色彩較為暗淡，而 JSON 提示詞生成影片的色調則明亮活潑。我無法決定哪個效果更好，所以留給大家自行判斷，請在下方評論區分享您的想法。
JSON格式提示詞技巧
- 結構很重要：請務必使用正確的 JSON 格式，並使用花括號、引號和逗號。
{
"request": "generate_image",
"style": "watercolor",
"details": "high"
}
- 使用巢狀物件進行復雜請求
{
"image": {
"type": "landscape",
"style": {
"medium": "oil painting",
"technique": "impasto"
}
}
}
- 包括樣式參考示例
{
"style_reference": {
"artist": "Van Gogh",
"period": "Post-Impressionism"
}
}
- 設定優先順序：
{
"priority_elements": {
"main_subject": "foreground castle",
"secondary": "mountain backdrop"
}
}
- 驗證你的提示詞：使用 JSONLint 等工具驗證你的 JSON
{
"fallbacks": {
"style": ["realism", "semi-realism"]
}
}
- 迭代最佳化：從基本結構入手，逐步新增細節，並維護 JSON 提示的版本控制。
小結
JSON 提示的真正威力不僅在於其結構，更在於其思維方式。與通常產生通用輸出的基本提示不同，JSON 要求精準，為 AI 提供清晰的執行框架，同時又不失創造力。細節之處見真章：更豐富的視覺效果、更智慧的邏輯，以及真正理解您設想的輸出。無論是動態網頁元素，還是富有深度和意圖的影像，JSON 提示不僅能滿足預期，還能展現 AI 的真正實力。
評論留言