OpenMontage Video Agent 提示词工程指南:从 Brief 到生产级提示词的完整写法(prompt-examples.md 详解)-CSDN博客

OpenMontage Video Agent 提示词工程指南：从 Brief 到生产级提示词的完整写法（prompt-examples.md 详解） 最新推荐文章于 2026-09-23 14:53:50 发布 原创 于 2026-09-07 15:51:12 发布 · 798 阅读 · OpenMontage Video Agent 提示词工程指南：从 Brief 到生产级提示词的完整写法（prompt-examples.md 详解） 【免费下载链接】OpenMontage World's first open-source, agentic video production system. 12 production pipelines, 100+ tools, 700+ agent skill and production-knowledge files. Turn your AI coding assistant into a full video production studio. 项目地址: https://gitcode.com/GitHub_Trending/op/OpenMontage 本文围绕 OpenMontage 仓库中 create-video 技能的核心参考文档 prompt-examples.md 展开，完整解析“输入简报 → 生产级提示词”的转换范式，并给出 5 套可直接套用的视频提示词模板（Tech News Briefing、Product Comparison、Strategy Presentation、Social Ad、Premium Report）。读完本文，你能掌握 HeyGen Video Agent 提示词的完整结构（FORMAT / TONE / AVATAR / STYLE / CRITICAL TEXT / SCENE / MUSIC / NARRATION），并理解每一节写法背后的底层原因，从而为代理式视频生产系统写出一次生成即接近成片的提示词。

1. 文档在 OpenMontage 中的定位

OpenMontage 是一个代理式（agentic）视频生产系统，仓库内 .agents/skills/ 目录存放着 700+ 个代理技能与生产知识文件。其中 create-video 技能 的职责是：给定一句文字需求，调用 HeyGen Video Agent 自动完成剧本、形象选择、画面、配音、节奏与字幕的全流程视频生成。技能文档明确了一条参考文件加载链： 文件 加载时机 prompt-optimizer.md 撰写提示词的核心工作流与规则 visual-styles.md 选择视觉风格（20 种命名风格完整规格） prompt-examples.md 从零写提示词时的完整生产示例 + 模板 video-agent.md Video Agent API 端点细节 也就是说，本文的指定文档 prompt-examples.md 处于“方法论（optimizer）→ 风格库（styles ）→ 示范与模板（本文档）”链条的最后一环：前两者告诉你“规则是什么”，本文档则给出“真正可以投喂给 API 的成品长什么样”。技能主文件 SKILL.md 中的 Default Workflow 也明确要求：写提示词前必须走 prompt -optimizer → visual-styles 的路径，再调用 mcp__heygen__generate_video_agent（或直连 POST /v1/video_agent/generate）。

2. 核心洞察：Video Agent 是一个“HTML 解释器”

在解读示例之前，先继承姊妹文档 prompt-optimizer.md 的关键论断，它是理解 prompt- examples .md 中所有写法的钥匙： Video Agent 是一个 HTML 解释器。它原生渲染版式、排版和结构化内容。B-roll（空镜）要用带动作动词的分层文字动效语言描述（“slams in”、“types on”、“counts up”），而不是布局规格（“upper-left, 48pt”）。 这条洞察直接决定了第 3 节完整示例中 B-roll 场景的写法——你会看到场景 2 到场景 5 全部采用 LAYER 1..5 分层结构，且每个元素都配有动作动词。同文档还总结了几条“反面模式”（What Doesn't Work），示例文档正是对这些模式的正面示范： 布局语言导致空画面：写 UPPER-LEFT: headline in 48pt Helvetica 这类屏幕坐标会产生空白/黑屏的 B-roll；正确写法是 "$95M" SLAMS in from left, white Impact 120pt, fills 40% of frame。 只写设计师名字不写规格：“Ikko Tanaka style” 对 Video Agent 没有意义，必须翻译成具体规则（每帧最多 3 种平面色、60% 留白、排版作为主视觉元素）。 把示例场景塞进提示词：风格库里完整的示例场景会干扰代理，只应引用风格的规则，而不是示例场景。 强行短 B-roll（≤5 秒）：过短的 B-roll 渲染不出内容，所有测试过的 5 秒 B-roll 视频都出现了空白/黑屏，应使用 10–15 秒。 内容写成清单而非故事：“这里有 5 条推文”这种平铺直叙会产出平淡的视频，必须先综合出论点：“X 正在发生，因为 Y——这是证据。”

3. 完整示例：从简报到生产提示词（Full Example）

prompt-examples.md 的主体是一个完整的“Brief → Production Prompt”案例，下面完整继承。 3.1 输入简报（Input Brief） Topic: Monthly company report for a SaaS startup Key data: $141M ARR (up from $54M), 1.85M signups (+28%), 3M paid videos/month Customer story: Creator built AI character, 2.5M followers, 20 min/video Challenge: Organic traffic volatile, -16% last week Duration: ~90 seconds Tone: Confident CEO, data-backed 简报只有 6 行：主题、关键数据、客户故事、挑战、时长、语气。这是典型的“素材散点”状态——它还不是提示词。 3.2 输出提示词（Output Prompt，完整） FORMAT: Bloomberg-style company report. 90 seconds. Fast-paced, data-dense. Record-breaking month. Proud but analytical. TONE: Confident, direct, data-backed. Highlights hit hard with numbers. Customer stories are the emotional core. Challenges are honest — no spin. AVATAR: Man in simple black crew-neck tee, standing in a modern glass-walled office at golden hour. Behind him, a wall-mounted display shows the company logo in soft blue glow. Monitor to his right shows a dashboard with upward-trending charts. Desk beside him: laptop, half-empty flat white, scattered sticky notes. Warm afternoon light through floor-to-ceiling windows, long shadows on polished concrete. Minimal, focused startup HQ. STYLE — SWISS PULSE (Müller-Brockmann): Grid-locked compositions. Black (#1a1a1a), white, electric blue (#0066FF), warm amber (#FF9500) for records. Helvetica Bold headlines, Regular labels. Numbers LARGE. Animated counters count up from 0. Diagonal compositions on accent moments. Grid wipe transitions. No dissolves. CRITICAL ON-SCREEN TEXT (display literally): • "1.85M SIGNUPS — +28% MoM" • "$2.12M NEW SUBSCRIPTION REVENUE" • "$54M → $141M ARR" • "2.5M FOLLOWERS" and "20 MIN / VIDEO" • Quote: "Use technology to serve the message, not distract from it." • "ORGANIC: 65% OF SUBS — VOLATILE" MUSIC: Upbeat electronic with a driving beat. Tycho meets Bloomberg opening theme. Builds through highlights, warms for customer story, softens for challenges, peaks on close. SCENE 1 — A-ROLL (8s) [Avatar center-frame, energetic, leaning slightly forward] VOICEOVER: "January was a record month. New highs across acquisition, revenue, and product velocity. Here's the full picture." Lower-third SLIDES in: "COMPANY NAME | JANUARY 2026" white on blue bar. Grid wipe. SCENE 2 — FULL SCREEN B-ROLL (12s) [NO AVATAR — motion graphic only] VOICEOVER: "One-point-eight-five million signups — twenty-eight percent month over month. Two-point-one-two million in new subscription revenue. Both all-time highs." LAYER 1: Dark #1a1a1a background with thin grid lines pulsing at 8% opacity. LAYER 2: "1.85M" SLAMS in from left, white Bold 140pt. "SIGNUPS" types on in electric blue 32pt uppercase. "+28% MoM" appears in amber. LAYER 3: Three stat cards CASCADE from top-right, staggered 0.3s: "$2.12M New Revenue" — "$3.4M Business ARR" — "$3M Pro ARR." Each number COUNTS UP from 0. LAYER 4: Bottom ticker scrolls: "Non-brand search +36% • Brand impressions 9.2M • Weekly subs +20.5%" LAYER 5: Grid lines RIPPLE outward on "1.85M" slam. Diagonal amber bar behind stat cards. Hard cut. SCENE 3 — FULL SCREEN B-ROLL (12s) [NO AVATAR — motion graphic only] VOICEOVER: "Zoom out. Twelve months ago — fifty-four million ARR. Today — one hundred forty-one million. Nearly three X in a single year." LAYER 1: Dark background, subtle grid scrolling upward. LAYER 2: Animated line chart DRAWS ITSELF left to right. Y-axis: $50M to $150M. Final point "$140.84M" glows amber and pulses. LAYER 3: Milestone annotations float in at key data点. LAYER 4: Second smaller chart below — "Paid Videos" 0.91M to 2.97M, same style. LAYER 5: Thin grid lines converge toward final data point. Scan line sweeps. Grid wipe. SCENE 4 — A-ROLL (8s) [Avatar center-frame, warm tone, genuine smile] VOICEOVER: "But the numbers only tell half the story. The other half is the people building on the platform." Lower-third: "Customer Spotlight" SCENE 5 — FULL SCREEN B-ROLL (12s) [NO AVATAR — warm palette] VOICEOVER: "An AI character built entirely on the platform. Twenty minutes per video. Two-point-five million Instagram followers. The creator's principle: use technology to serve the message, not distract from it." LAYER 1: Dark background with warm amber grid lines at low opacity. LAYER 2: "CHARACTER NAME" in large white, center-top, 80pt. LAYER 3: Stats cascade from right: "2.5M Followers" COUNTS UP in amber — "20 min/video" — "7x Faster." Each a glowing node. LAYER 4: Quote card SLIDES UP: "Use technology to serve the message, not distract from it." Types on word by word. LAYER 5: Warm light bloom. Grid lines soften into curved arcs. Grid wipe. SCENE 6 — A-ROLL (10s) [Avatar center-frame, serious/candid] VOICEOVER: "Now the honest part. Organic drives sixty-five percent of subscriptions and it's volatile. Non-brand traffic dropped sixteen percent last week. We've rebuilt attribution and we're investing in SEO." Lower-third: "Challenges" SCENE 7 — A-ROLL (7s) [Avatar center-frame, energy lifts, direct eye contact] VOICEOVER: "Fifty-four million to one-forty-one in twelve months. Three million paid videos a month. January set the bar — now we raise it." End card: Logo centered, blue glow fade-in. Grid lines converge. Music peaks. NARRATION STYLE: CEO energy — conviction backed by data. Fast on highlights. Warm on customer stories. Candid on challenges. Close with forward momentum. 3.3 逐节拆解：为什么这样写 对照 prompt-optimizer.md 中的 Prompt Anatomy，生产级提示词固定为八段结构，上文示例每一段都给出了可执行的写法： FORMAT——视频类型 + 时长 + 能量基线。示例写作 Bloomberg-style company report. 90 seconds. Fast-paced, data-dense.，并用 Record-breaking month. Proud but analytical. 一句话锚定情绪。 TONE——情感寄存器 + 参照系。注意简报中的 Tone: Confident CEO, data-backed 被展开成了三条规则（数字要硬、客户故事是情感核心、挑战不粉饰），把抽象语气翻译成了可执行的播报策略。 AVATAR——60–100 词的角色 + 场景设计。optimizer 强调“avatar 不是固定大头照，要为每条视频像电影角色一样设计（服装设计师 + 布景师思维）”。示例严格遵守：服装（黑色圆领 T 恤）、品牌元素（墙挂屏上发蓝光的公司 logo）、内容相关道具（右侧显示器是上涨曲线 dashboard、桌上散落便利贴）、光线（黄金时刻暖光、抛光水磨石上的长影）。这些正对应 optimizer 的 Avatar 模板：[Clothing][Setting][Monitors/desk][Lighting][Mood]，要求 3+ 个内容相关道具、品牌元素可见。 STYLE——命名风格 + 具体规则。STYLE — SWISS PULSE (Müller-Brockmann) 直接引用了 visual-styles.md 风格库中第 6 号风格（黑 #1a1a1a / 白 / 电光蓝 #0066FF、Helvetica Bold、网格锁定、计数器从 0 累加、网格擦除转场、禁用叠化），并追加了“破纪录时刻用暖琥珀色 #FF9500”的内容定制。这正是 optimizer 的警告“只给设计师名字没有用”的正确解法：给名字（便于风格库对齐）+ 给具体规则（可执行）。 CRITICAL ON-SCREEN TEXT——必须逐字呈现的字符串清单。简报里的 6 个数据点被转写为大写、标准化的屏幕文本（"1.85M SIGNUPS — +28% MoM"、"$54M → $141M ARR" 等）。optimizer 说明：没有这一节，代理可能会总结、四舍五入或改写引语。这是数据类视频准确性的第一保障。 MUSIC——类型 + 参照艺术家 + 能量弧线。示例的 Tycho meets Bloomberg opening theme 加四段能量变化（highlights 推进 → 客户故事转暖 → 挑战段放缓 → 结尾冲顶），把配乐从“随便放个 BGM”变成了有叙事结构的声轨。 SCENE 1–7——逐场景剧本，详见 3.4。 NARRATION STYLE——配音的整体演绎方式：CEO energy — conviction backed by data. Fast on highlights. Warm on customer stories. Candid on challenges. Close with forward momentum. 3.4 场景结构：类型轮换与五层视觉系统 示例的 7 个场景完整演示了 optimizer 定义的三种场景类型（A-ROLL / FULL SCREEN B-ROLL / A-ROLL + OVERLAY）与强制轮换规则（同一类型不得连续 3 个以上、至少 2 个纯 B-roll 场景、每个场景包括 B-roll 都必须有 VOICEOVER）： 场景 类型 时长 作用 SCENE 1 A-ROLL 8s Hook + 下三分之一标题条 + 网格擦除 SCENE 2 FULL SCREEN B-ROLL 12s 注册量/营收数据，L1–L5 五层动效 SCENE 3 FULL SCREEN B-ROLL 12s ARR 增长曲线自绘动画 SCENE 4 A-ROLL 8s 从数据过渡到“人” SCENE 5 FULL SCREEN B-ROLL 12s 客户故事 + 逐字打字引语卡 SCENE 6 A-ROLL 10s 直面挑战（有机流量 -16%） SCENE 7 A-ROLL 7s 收尾 + End card（logo 蓝光淡入，音乐冲顶） 每个 B-roll 场景都严格执行了 optimizer 的“视觉分层系统”：L1 背景（脉冲网格）、L2 主角内容（巨大的 “1.85M”）、L3 支撑数据（错峰 0.3s 级联的三张数据卡）、L4 信息条（底部 ticker 滚动）、L5 特效（网格涟漪、斜切琥珀色条）。规则是：每个 B-roll 至少 4 层、每个元素都必须动（Every element must MOVE）。 动作文法（Motion Vocabulary） 在示例中被分三档使用：高能动词（SLAMS / CRASHES / PUNCHES / STAMPS / SHATTERS）用于数据冲击时刻，如 “1.85M” SLAMS in from left, white Bold 140pt；中能量动词（CASCADE / SLIDES / DRAWS / FILLS）用于常规数据呈现，如 Three stat cards C...