Video PRD Dual Agent

Core Positioning

video-prd-dual-agent is a controller skill for mandatory real multi-agent PRD production.

It does not simulate an internal team. The main agent is the Controller / Integrator: it reads the materials, dispatches visible role-bounded sub-agents, collects their artifacts, integrates the final PRD, and reports execution evidence.

Forbidden:

• silently simulating Agent A / Agent B / Scene Planner / Scene Director / Motion Agent / Text Agent / PRD Reviewer inside one hidden draft
• claiming multi-agent execution without listing dispatched agents and returned artifacts
• asking the user for step-by-step decisions when the provided materials and defaults are enough
• producing a PRD that lacks Scene, Shot, Motion, subtitle, prompt, or risk-review sections

Invoking this skill counts as an explicit user request for role-bounded sub-agent work.

Mandatory Real Multi-Agent Execution

Before generating any PRD content, the main agent must:

1. Load this SKILL.md and the required reference checklist.
2. Read all user-provided text materials completely, including SRT files.
3. Inspect all user-provided local images, screenshots, charts, and media that are relevant to the PRD.
4. Dispatch visible sub-agents for all mandatory roles:
   • Agent A
   • Agent B
   • Scene Planner
   • Scene Director
   • Motion Agent
   • Text Agent
   • PRD Reviewer
5. Collect one concrete artifact from each role.
6. Assemble PRD v1 only after role artifacts are returned.
7. Send PRD v1 to PRD Reviewer.
8. If PRD Reviewer returns NOT OK, route mandatory fixes back to the owning role or fix them locally when the ownership is purely integrative, then repeat reviewer validation until OK or until the blocker is explicit.

If visible sub-agent capability is unavailable, do not silently degrade into single-agent simulation. Return exactly:

Blocked: visible sub-agent capability unavailable; this skill forbids silent single-agent simulation.

If any mandatory role cannot be dispatched or fails to return a concrete artifact, do not continue as if the role succeeded. Return exactly:

Blocked: mandatory role artifact unavailable: <role-name>. This skill forbids completing the PRD without all required role artifacts.

Execution Evidence

The final response must include a short execution evidence block before or after the PRD location:

• Selected agents
• One-line artifact returned by each agent
• PRD Reviewer verdict
• Final PRD file path, when a file is created

Hide chain-of-thought and do not paste long agent transcripts unless the user explicitly asks for them. Do not hide the fact that agents were dispatched.

Initial Intake

First identify and analyze everything the user provided:

• SRT, narration script, voiceover draft, subtitles, outline, or story text
• screenshots, charts, data images, product images, brand visuals, style frames, or reference images
• product, brand, audience, platform, duration, aspect ratio, distribution context, and risk constraints
• existing local files mentioned in the thread, such as generated charts, reports, images, data files, or prior scripts

If there are images or local media, inspect them before writing the PRD. If there is text, read it completely before extracting the topic, structure, emotional arc, key claims, risk-sensitive statements, and phrases suitable for on-screen emphasis.

For detailed checkpoints, read references/video-prd-checklist.md before dispatching PRD roles.

Default Assumptions

Use these defaults only when the user does not explicitly specify a choice and the materials do not imply another answer.

• Default platform: Douyin / WeChat Video Account / Xiaohongshu short-video distribution.
• Default aspect ratio: 9:16 vertical, 1080x1920.
• If charts, dashboards, desktop UI, or 16:9 assets are central, use a 16:9 master version and note a 9:16 derivative when useful.
• If the user provides SRT, use the SRT total duration and timestamps as the timing baseline.
• If the user provides only narration, estimate duration from Chinese narration speed and split by semantic paragraphs.
• Default style: high information density, fast rhythm, strong opening hook, data-driven presentation, clear subtitles, highlighted keywords and numbers.
• The first 3 seconds must contain at least one strong conclusion, question, key statistic, conflict point, or suspense point.
• Default final target: HyperFrame / HyperFrames video generation workflow.

Role Contracts

The main agent is the Controller / Integrator. It is not allowed to impersonate the role outputs below.

Agent A: Requirements Designer

Input:

• user request
• complete text/SRT summary
• visual asset inventory
• chart/data facts
• known risk constraints

Output artifact:

• decision memo covering objective, audience, platform, ratio, duration, style, opening hook, narrative structure, chart/data strategy, and risk boundary
• concrete questions for Agent B to agree with, modify, or reject

Agent B: Co-Designer

Input:

• Agent A decision memo
• same material summary and risk constraints

Output artifact:

• review memo
• for each decision: agreed, modified, or disagreed
• executable alternative when modifying or disagreeing
• list of decisions that must be reflected in the final PRD

Collaboration rule:

• Agent A proposes; Agent B responds.
• If Agent B modifies or disagrees on core decisions, send the response back to Agent A once for reconciliation.
• Limit the A/B loop to two rounds unless the conflict would make the PRD unusable.

Scene Planner

Input:

• reconciled Agent A/B decision memo
• SRT or narration timeline
• source material map

Output artifact:

• Scene overview table
• per-Scene time range, narration segment, narrative purpose, core message, recommended visual type, source material, emotional rhythm

Scene rules:

• Split by content logic, not mechanical equal duration.
• Opening hook must be its own Scene.
• Align timing with SRT timestamps when SRT exists.

Scene Director

Input:

• Scene Planner artifact
• visual asset inventory
• HyperFrame target constraints

Output artifact:

• Shot-by-Shot storyboard table
• per-Shot time range, narration, visual content, subject elements, background elements, camera movement, transition, required assets, HyperFrame-oriented prompt, notes

Shot rules:

• Each Scene must include at least one Shot.
• Complex Scenes should be split into multiple Shots.
• Do not write only "show chart" or "show image"; describe entry motion, highlighted region, viewer focus, and transition.

Motion Agent

Input:

• Scene Planner artifact
• Scene Director storyboard

Output artifact:

• Motion timeline table
• per-Shot animation timing, objects, actions, animation style, duration, intensity, and purpose

Motion rules:

• The first 3 seconds need clear rhythm design.
• Important numbers may enlarge, outline, glow, or sweep-light, but avoid empty spectacle.
• Chart content should reveal progressively rather than appearing all at once.
• Motion must serve information clarity.

Text Agent

Input:

• SRT or narration
• Scene and Shot artifacts
• risk constraints

Output artifact:

• subtitle and keyword table
• main on-screen title and supporting text rules
• highlighted words and number rules
• banned-word check
• final risk wording

Text rules:

• No more than 2 subtitle lines per screen.
• Prefer no more than 14 Chinese characters per line for vertical video; for 16:9 chart videos, use readable line breaks and avoid chart overlap.
• Important numbers should be isolated or enlarged when useful.
• Subtitles must match the narration meaning.
• Screen titles may compress wording but must not distort the SRT/narration.

PRD Reviewer

Input:

• PRD v1 assembled by the Controller / Integrator
• all role artifacts
• source material summary

Output artifact:

• review verdict ending with exactly OK or NOT OK
• if NOT OK, list only mandatory fixes with section references

Reviewer must check:

• completeness
• executable Scene design
• concrete Shot design
• concrete Motion rules
• clear subtitle rules
• consistency with user materials
• HyperFrame / HyperFrames suitability
• finance, medical, marketing, copyright, or other domain risk controls when relevant
• absence of fabricated data
• absence of forbidden expressions

Finance And Investment Risk Rules

Activate finance risk review when the material involves stocks, funds, ETFs, futures, macroeconomics, financial reports, industry investing, trading strategies, individual stock analysis, price movement predictions, position sizing, or return statistics.

Finance content should frame results as historical observation, product demonstration, methodology explanation, and learning exchange. It must not be framed as investment advice.

Do not output concrete trading execution advice, including:

• buy-in price ranges
• take-profit or stop-loss levels
• position-size percentages or allocation ratios
• leverage multiples
• intraday follow-trade instructions
• packaging backtest or historical win rate as future return expectation

Forbidden or must-be-softened expressions include:

• 必涨
• 稳赚
• 保证收益
• 无风险
• 无脑买入
• 满仓
• 梭哈
• 确定性机会
• 明确买入
• 明确卖出
• 目标价一定达到
• 100% 胜率
• 包赚
• 翻倍确定
• 暴富机会

Prefer safer wording such as:

• 历史复盘
• 方法演示
• 数据观察
• 学习交流
• 不构成投资建议
• 需要继续验证
• 历史数据不代表未来
• 理论最优是事后观察指标，不代表实盘可达到

If user material contains high-risk language, preserve the factual context but rewrite screen text into non-actionable analysis, risk observation, method explanation, or educational framing. SRT subtitles may follow the narration, but large screen text must be risk-controlled.

SRT And Material Handling Rules

If the user provides SRT:

• read SRT bytes with encoding detection when needed
• preserve timestamps
• align Scene and Shot timing with SRT as much as possible
• use SRT text as narration ground truth unless the user provides a newer script
• do not omit important narration information
• merge subtitle items into larger Scenes while preserving logical order

If terminal output shows garbled text, verify file bytes and encoding before concluding that the SRT is corrupt.

If multiple materials conflict, resolve with these priorities:

1. Timing: SRT timestamps > narration timing estimate
2. Subtitle meaning: newest user-provided script > SRT text > older conversation draft
3. Data: chart or screenshot original text and numbers > narration summary
4. Visual style: provided brand/material visuals > generic defaults

Record any assumption in the internal decision summary.

PRD Output Structure

Use Chinese. Prefer concise but complete Markdown.

# AI 视频生成 PRD

## 0. 多 Agent 执行摘要

## 1. 项目目标

## 2. 输入素材分析

## 3. 需求决策总表

## 4. 视频定位

## 5. Scene 总览

## 6. Scene-by-Scene 结构

## 7. Shot-by-Shot 分镜 PRD

## 8. Motion 动画时间轴

## 9. 字幕与关键词高亮规则

## 10. 视觉语言

## 11. 统一视觉规范

## 12. AI 视频生成总提示词

## 13. Negative Prompt / 禁止事项

## 14. 风险与合规检查

## 15. 生成后检查清单

Section Requirements

0. 多 Agent 执行摘要

Include:

• selected agents
• one-line artifact returned by each agent
• PRD Reviewer verdict
• major assumptions

Do not include long transcripts or hidden reasoning.

1. 项目目标

Explain:

• what problem the video solves
• what the audience should understand
• what the audience should remember
• what the distribution goal is

2. 输入素材分析

Organize:

• SRT / narration script
• screenshots
• charts
• product materials
• brand materials
• statistics
• other assets

Analyze how each material should be used. If a material type is not provided, write 未提供; do not invent it.

3. 需求决策总表

Use:

决策项	最终决策	来源 Agent / 工件	理由	风险/备注

Include platform, aspect ratio, duration, style, opening hook, narrative structure, subtitle rules, chart usage, animation intensity, risk reminder, and AI generation method.

5. Scene 总览

Use:

Scene	时间范围	叙事功能	核心信息	推荐画面

Opening hook must be listed separately.

7. Shot-by-Shot 分镜 PRD

Use:

Shot	时间范围	对应口播	画面内容	主体元素	背景元素	镜头运动	转场	所需素材	AI 视频生成 Prompt	备注

Each Shot must include a HyperFrame-oriented prompt.

8. Motion 动画时间轴

Use:

Scene	Shot	时间点	对象	动作	动画方式	持续时间	强度	目的

Every Motion row must link to a specific Scene and Shot.

9. 字幕与关键词高亮规则

Include:

• subtitle position
• subtitle font size
• max characters per screen
• highlighted keywords
• highlight style
• large on-screen text rules
• risk reminder rules
• forbidden words

If SRT exists, output:

时间范围	字幕	屏幕大字	高亮词	样式

12. AI 视频生成总提示词

Output one global HyperFrame prompt that can be directly used by the HyperFrame / HyperFrames video generation workflow.

Must include:

• video style
• aspect ratio
• camera language
• motion style
• subtitle style
• chart / screenshot display style
• pacing
• risk wording requirements
• instruction not to fabricate data
• instruction to follow the Scene, Shot, and Motion timeline exactly

14. 风险与合规检查

If finance content is involved, this section is mandatory and must include:

• whether it involves investment advice
• whether it involves individual stocks / funds / ETFs
• whether it includes return-promise risk
• whether it includes exaggerated language
• whether data source is unclear
• high-risk expressions that were replaced
• final risk reminder wording

If not finance content, keep the section and write the relevant non-finance risk check.

Completion Checklist For Skill Edits

When editing this skill, verify:

1. SKILL.md contains no garbled Chinese.
2. It explicitly forbids silent single-agent simulation.
3. It requires visible sub-agent dispatch.
4. Every mandatory role has an output artifact contract.
5. PRD Reviewer has an OK / NOT OK gate.
6. Final output requires execution evidence.
7. HyperFrame / HyperFrames remains the primary output target.
8. Finance risk-control rules are included.
9. SRT handling requires reading actual text, not only timestamps.
10. PRD includes executable Scene, Shot, Motion, subtitle, prompt, and risk sections.