fpv-immersive-video-prompting
Route-first FPV AI video prompting skill for Seedance, Kling, Runway, Veo, GPT Image asset packs, numbered stops, and red-line path control.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/zhouwei713--fpv-immersive-video-prompting)HTML
<a href="/listings/zhouwei713--fpv-immersive-video-prompting"><img src="/api/listings/zhouwei713--fpv-immersive-video-prompting/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
方式一:安装到 Hermes
git clone https://github.com/zhouwei713/fpv-immersive-video-prompting.git \ ~/.hermes/skills/creative/fpv-immersive-video-prompting
重启 Hermes,或开启新会话后即可使用:
fpv-immersive-video-prompting
方式二:作为通用 Skill 使用
如果你不使用 Hermes,也可以直接把 `SKILL.md` 放进 Claude、Codex、Cursor、OpenCode 或其他支持 Skill / long prompt 的 Agent 环境里。
方式三:只当提示词方法论参考
直接阅读:
SKILL.md skill/references/gpt-image-asset-packs.md skill/references/session-patterns.md
---
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 235 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- ---
- name: fpv-immersive-video-prompting
- description: Use when the user wants to turn a static scene, character images, aerial map, drawn path, or route-control image into an immersive first-person FPV AI video prompt, especially Seedance/Kling/Runway/Veo style one-shot videos with numbered stop markers, red-line path control, world-map flythroughs, camera route planning, variable character counts, non-human POVs such as drones, pets, robot vacuums, character references, timed interactions, dialogue, spatial audio, and negative constraints.
- version: 1.0.0
- author: Hermes Agent
- license: MIT
- metadata:
- hermes:
- tags: [video, ai-video, fpv, prompting, camera-route, character-consistency]
- related_skills: [gpt-image-2-prompting]
- ---
- # FPV Immersive Video Prompting
- ## Overview
- This skill turns a scene idea plus optional image references into a directed FPV AI-video prompt. The target output should feel like the viewer is moving through a playable scene, not watching a generic beauty shot.
- Use this for image-to-video workflows where the user wants:
- - one-shot first-person walkthroughs
- - numbered route stops or camera planning
- - characters reacting as the camera approaches
- - human or non-human POVs: guest, pet cat, drone, robot vacuum, bird, vehicle, object, spirit
- - variable numbers of characters or targets
- - strong identity and route consistency
- ## Core Principle
- Default to Chinese for all generated prompts, including GPT Image asset prompts, route-control image prompts, video prompts, negative prompts, and delivery notes, unless the user explicitly asks for English or a target tool requires English. Keep technical terms such as FPV, Seedance, Kling, Runway, Veo, one-shot, prompt, and path control in English when they are clearer.
- Write the output as a small playable scene specification:
- 1. Who or what is the camera?
- 2. Where does it start?
- 3. How many main targets are there exactly?
- 4. What numbered order does it visit them in?
- 5. How does the camera physically move?
- 6. What happens at each stop?
- 7. What must stay consistent?
- 8. What must never appear?
- ## Default Asset Workflow
- When the user needs GPT Image / GPT-Image-2 assets, request a multi-image asset pack rather than one crowded contact sheet. Treat requests for “生图 prompt”, “首帧图”, “参考图”, “素材包”, “用 GPT Image 生成”, or any image-prep step as an asset-pack request, not as a single first-frame prompt.
- For any case that needs multiple images, output one batch-generation prompt that asks GPT Image / ChatGPT to generate multiple separate images in a single response. Do not split the asset pack into many independent per-image prompts unless the user explicitly asks for that. The prompt must say: “请一次性生成 [X] 张独立图片,不要生成拼图、九宫格、contact sheet 或一张图里塞多个画面。”
- For close-interaction scenes with N main people/targets, always output the complete asset pack unless the user explicitly asks for only one image:
- 1. First-frame scene image with small numbered stop markers only: 1, 2, 3, ... N
- 2. N separate character/reference images, one for each main person/target
- 3. Optional clean first frame without numbers for safer image-to-video input
- Before returning any image prompt, run this count check: total images = 1 route/numbered first frame + N character references + optional 1 clean first frame. The video prompt must refer to these images by role, e.g. “图片 1 is route planning, 图片 2-4 are character references, 图片 5 is clean first frame.”
- Default image sets:
- Close-interaction / character mode:
- 1. First-frame scene image with small numbered stop markers only: 1, 2, 3, ...
- 2. One character/reference image for each main target
- 3. Optional clean first frame without numbers for safer image-to-video input
- World-route / path-control mode:
- 1. Aerial route-control map or wide first frame with one clear continuous red path from start to destination
- 2. Optional clean world/scene reference without the red path
- 3. Optional landmark or character references only if specific destinations need consistent appearance
- Prefer numbered stop markers for close-range character interactions, indoor scenes, crowded social scenes, and exact target-count workflows. GPT Image often struggles to draw one continuous physically coherent route through furniture, people, walls, railings, or water; a bad red line can mislead the video model. Numbered stops are easier to generate correctly, and the video prompt can define the movement between them.
- Use red-line path control for large-scale route-shaped scenes: aerial world maps, fantasy continent journeys, city-to-landmark flythroughs, racing lines, canyon/drone routes, open-world game trailers, and Seedance 2.0 path-control demos. In that mode, the route image is a planning artifact; the final video must remove all red lines, arrows, annotations, labels, and map-view appearance.
- ### Red-Line Path Control Rules
- Use this mode when the user asks for Seedance 2.0 path control, a drawn route, an aerial map flythrough, a game-world traversal, a racing line, or a large-scale drone/bird/spirit FPV journey.
- For the route-control image prompt:
- - Use a high-resolution 16:9 aerial terrain map, world map, tactical map, or wide route-planning scene.
- - Draw one clear continuous red route line, optionally with a subtle arrow, from start to destination.
- - Make the route physically plausible through visible corridors: roads, valleys, rivers, city gates, bridges, canyons, rooftops, airspace, tunnels, coastlines, or ridgelines.
- - Build 4-6 visually distinct route segments so the video has progression: peaceful start, transition zone, landmark/city, danger zone, final destination.
- - Keep the red line clean and legible. Avoid multiple competing paths unless the user explicitly wants branching routes.
- For the video prompt:
- - Say the uploaded image is a route-planning map/image, not the final look.
- - Say the red line/arrow/annotations are only camera-path controls and must be completely removed.
- - State the camera must strictly follow the drawn route geometry from start to destination.
- - Use invisible first-person drone, bird, spirit, vehicle, or mounted-camera POV unless the route is truly walkable.
- - Add timeline segments by geography, not by character stop.
- - Include camera language: continuous forward motion, natural banking, close passes, altitude changes, foreground parallax, progressive acceleration, smooth horizon control.
- - Avoid map view, visible red lines, annotations, teleporting, reverse motion, jump cuts, visible drone, guide characters, flat terrain, deformed landmarks, flickering structures.
- Use numbered stops instead when the video depends on close character interactions, exact character count, indoor navigation, or object-level continuity.
- ### Numbered Stop Marker Rules
- For the first-frame image prompt:
- - Do not draw red route lines, connecting lines, or arrows by default.
- - Use small numbered markers near each target or on the floor beside each target.
- - Markers must be ordered by reachable spatial depth, not randomly scattered.
- - Marker 1 should usually be closest to camera; later markers move deeper into the scene.
- - Each next stop must be physically reachable from the previous one without crossing walls, water, furniture, railings, people, or other obstacles.
- - The scene must be designed around a walkable/glideable/flyable route before placing characters.
- Only use a continuous red route line if the user explicitly asks for it or the specific workflow handles route lines reliably.
- ## Variable Count Handling
- If the user specifies a number of people/targets, build the whole asset plan, route, timeline, and video prompt around exactly that number. Do not default back to five.
- Recommended target counts:
- | Duration | Best count | Handling |
- | --- | --- | --- |
- | 5-8s | 1-3 targets | one clear approach, few actions |
- | 8-12s | 3-4 targets | short interactions, one final reveal |
- | 12-15s | 4-5 targets | ideal for sequential stops |
- | 15-25s | 5-8 targets | shorter beats and group transitions |
- | 25s+ | 8+ targets | split into groups/zones |
- If the user says “a group” without a number, choose a practical default based on duration, usually 4-5 main interaction targets. Background extras may exist only if requested and must remain secondary.
- Always write:
- “exactly [N] main characters/targets” and “do not add or remove main targets.”
- ## POV Identity Selection
- FPV does not always mean normal human eye level. If the user specifies the POV, honor it. If not, choose a POV that makes the scene more coherent or interesting.
- | Scene / intent | Good POV choices | Movement cues |
- | --- | --- | --- |
- | palace, mansion, apartment, party | invited guest, butler, courier, pet cat | walking sway, room turns, hand/paw glimpses |
- | modern living room, cafe, cozy home | pet cat, robot vacuum, visitor | low-angle gliding, furniture legs, paws/tail or wheel hum |
- | outdoor landscape, city, battlefield | drone, bird, kite, spirit | aerial arcs, height changes, wind, no footstep bob |
- | museum, exhibition, showroom | visitor, guide robot, security camera | slow walkthrough, display pauses, precise turns |
- | horror / mystery | child, flashlight holder, CCTV, ghost | low height, hesitant movement, limited view |
- | car/train/boat/aircraft | passenger, driver, mounted camera | constrained path, window motion, engine/road/water sounds |
- POV constraints:
- - Human: eye-level height, walking bob, breathing, hands/sleeves may appear.
- - Pet cat/dog: low height, curious head turns, paws/tail may appear, no human hand gestures.
- - Drone: smooth continuous flight, altitude changes, banking turns, propeller/wind sound, no walking bob.
- - Robot vacuum/small robot: very low height, slow floor gliding, furniture legs loom large, mechanical hum, cannot climb stairs or jump.
- - Bird/insect/spirit: floating arcs or gliding, unusual height, but still continuous and motivated.
- - Object POV: movement must follow the object’s plausible motion unless the concept is surreal.
- Always enforce physical limits. A cat cannot float over a table; a drone cannot pass through walls; a robot vacuum cannot jump onto a sofa; a human cannot squeeze through furniture.
- ## Video Prompt Structure
- Default structure:
- 1. Input usage: first frame, numbered route reference, character references, or red-line route-control image
- 2. Specs: aspect ratio, duration, one-shot
- 3. POV identity and physical movement
- 4. Route order: numbered target stops for close interaction, or geography segments for red-line path control
- 5. Compact timeline with one beat per target or one beat per route segment
- 6. Environment motion and sound design if useful
- 7. Consistency and avoid list
- Keep the final video prompt directly copyable. Do not impose a fixed character limit unless the user asks for one.
- ## Template: Video Prompt
- ```text
- 使用上传图片作为首帧、编号路线参考和 [N] 个角色/目标外观参考,生成一段 [比例]、[时长] 秒、一镜到底的 [场景类型] [POV身份] FPV 视频。首帧中的编号 [1..N] 只作为镜头停靠顺序参考,不要出现在最终画面里。最终视频不要出现数字、编号、路线、箭头、文字标签或 UI。
- 观众/摄像机是一位/一台 [POV身份],从 [起点] 出发,严格按编号顺序移动:[起点] → 1 [目标1] → 2 [目标2] → ... → [终点/回望]。本视频包含 exactly [N] 个主要人物/目标:[列出名称],不要增加或减少主目标。角色脸、服装、位置、身份保持参考图一致。
- 镜头运动必须符合 [POV身份] 的物理限制:[低机位/步行/飞行/滑行/车载等具体运动];移动连续,有真实加速减速,不能瞬移、跳切、穿过障碍物或变成其他 POV。
- 时间轴:
- 0.00-[入口时间]:[建立起点和环境动态]
- [时间]-[时间]:移动到 1,[目标1动作/台词/反应]
- [时间]-[时间]:移动到 2,[目标2动作/台词/反应]
- ...
- [结尾时间]:到达 [终点] 或回望全景,形成空间层次。
- 环境动态:[水波/窗光/咖啡蒸汽/灯光/风/反光/布料/屏幕等]。
- 避免:编号残留、路线/箭头、换脸、同脸、身份漂移、主目标数量变化、穿模、跳切、瞬移、错误 POV、过度抖动、场景漂移、低质、畸形、水印。
- ```
- ## Common Pitfalls
- 1. Asking GPT Image for a continuous red route line by default.
- - Prefer numbered stop markers. Continuous lines often become disconnected, cross obstacles, or contaminate the video.
- 2. Letting numbered stops scatter randomly.
- - First define a walkable/glideable/flyable spatial route, then place numbered stops along it.
- 3. Ignoring the user’s specified count.
- - Use exactly the requested number in assets, route, timeline, and constraints.
- 4. Treating every POV like a human.
- - Robot vacuums need floor gliding; cats need low curious movement; drones need flight; humans can show hands and walking sway.
- 5. Giving every person full dialogue in high-count scenes.
- - For more than 6 targets, use groups, gestures, glances, or quick beats.
- 6. Forgetting to remove markers in the final video.
- - Always say numbers/markers are planning references only and must not appear in final output.
- 7. Forgetting that red-line path control and numbered stops solve different problems.
- - Numbered stops are better for characters and interiors; red-line route control is better for world-scale flythroughs, racing lines, and map-to-world journeys.
- ## Verification Checklist
- Before returning the prompt, check:
- - [ ] Aspect ratio and duration are specified
- - [ ] First frame / route reference / character references are assigned
- - [ ] Main target count is explicit and matches the user request
- - [ ] POV identity is explicit or reasonably inferred
- - [ ] POV movement physics match the identity
- - [ ] Route order is explicit and physically continuous
- - [ ] Timeline covers the full duration and fits the target count
- - [ ] Each target has an action/reaction, or group beats for high-count scenes
- - [ ] Character/object consistency constraints are repeated
- - [ ] Environment secondary motion is included where useful
- - [ ] Negative constraints mention no route lines/arrows/numbers/markers in final video
- - [ ] Negative constraints mention no added/removed main targets
- - [ ] Red-line path-control mode is used only when the scene is route-shaped and large-scale, or the user explicitly requests it
- - [ ] Generated prompts are in Chinese by default unless the user explicitly requested English or a target tool requires English
- - [ ] Final prompt is directly copyable
- ## References
- Read `references/session-patterns.md` for examples and session-specific lessons from the initial development of this workflow.
- Read `references/mayz-seedance-world-route-case.md` for the Seedance 2.0 red-line world-route pattern: aerial map control image, strict drawn route geometry, map-to-world flythrough, biome progression, speed-run/racing-line mode, and theme-park ride style journeys.
- Read `references/public-article-angle.md` when explaining this workflow publicly in a WeChat article, X thread, tutorial intro, or demo write-up. It captures the session framing: FPV prompts are action trajectories, not just visual descriptions; numbered stops vs red-line path control; space-first design; POV physics; count-duration tradeoffs.
- Read `references/cafe-cat-numbered-stops-case.md` when the user asks for cafe, cozy indoor, pet-cat POV, or low-angle numbered-stop interaction scenes. It captures a 15-second / 3-person cafe route pattern and the cat-specific movement constraints that prevent drone-like motion.
README.md · README · 277 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- <div align="center">
- # FPV 运镜导演.skill
- > *把 AI 视频提示词,从“画面描述”升级成“行动轨迹设计”。*
- [](LICENSE)
- [](https://agentskills.io)
- [](#安装)
- [](#适合什么场景)
- <br>
- **一个专门为 FPV / image-to-video 设计的 AI 视频提示词 Skill。**
- 它不只帮你写“电影感、浅景深、光影高级”。
- 它会先设计镜头是谁,从哪里出发,按什么顺序经过哪些人,怎么绕过障碍物,最后停在哪里。
- <br>
- [看效果](#效果示例) · [安装](#安装) · [它解决什么](#它解决什么) · [工作原理](#工作原理) · [资产包](#gpt-image-资产包) · [English](README_EN.md)
- </div>
- ---
- ## 效果示例
- ### 咖啡厅猫咪视角
- 用户输入:
- ```text
- 帮我做一个咖啡厅里 3 个人依次互动的 FPV 视频提示词,15 秒,猫咪视角
- ```
- Skill 会先判断这不是红线路径场景,而是近距离人物互动。它会生成一套完整资产包,而不是只给一张首帧图。
- ```text
- 图片 1:咖啡厅猫咪低机位首帧,带 1、2、3 小编号停靠点
- 图片 2:靠窗女生独立人物参考图
- 图片 3:吧台咖啡师独立人物参考图
- 图片 4:角落圆桌男生独立人物参考图
- 图片 5:干净版首帧,去掉编号,用作真正视频首帧
- ```
- 然后输出视频 prompt:
- ```text
- 观众是一只在咖啡厅里自由走动的猫咪,镜头保持接近地面的低机位,
- 从咖啡厅入口旁的地垫开始,严格按编号顺序移动:
- 入口地垫 → 1 靠窗座位的女生 → 2 吧台前的咖啡师 → 3 角落圆桌旁的男生 → 窗边阳光下停住。
- 全片包含 exactly 3 个主要人物,不要增加或减少主目标。
- 镜头运动必须符合猫咪的身体限制,能看到桌腿、椅腿、地面纹理、人的鞋子,
- 可以短暂停顿、好奇转头、绕开障碍物,但不能飞,不能跳上吧台,不能穿过桌椅和人腿。
- ```
- 这类提示词的重点不是“咖啡厅很漂亮”,而是让视频模型知道猫到底怎么走。
- ---
- ### 世界地图飞行
- 用户输入:
- ```text
- 我想做一个 Seedance 2.0 的世界地图飞行,从雪原穿过峡谷、王城,最后到火山。
- ```
- Skill 会切换到红线路径控制模式。
- ```text
- 图片 1:16:9 奇幻大陆航拍路线规划图,一条连续红线从雪原出发,经过峡谷和王城,抵达火山口
- 图片 2:可选干净世界参考图,去掉红线
- ```
- 视频 prompt 会明确说明:
- ```text
- 红色路线只作为摄像机路径控制,不是最终画面内容。
- 最终视频不要出现红线、箭头、地图标注、文字标签、UI 或俯视地图感。
- 镜头必须严格沿红线几何飞行,有自然 banking、贴近地形掠过、穿越地标时的前景视差和稳定地平线。
- ```
- 同样是 FPV,一个是猫在咖啡厅里走,一个是无人机穿越大陆。两种场景不能用同一套提示词。
- ---
- ## 它解决什么
- 很多 AI 视频 prompt 看起来很完整,实际一生成就翻车。
- 常见问题是:
- - 镜头突然瞬移,空间断了
- - 人物数量一会儿多一会儿少
- - 首帧里的编号、红线、箭头残留在成片里
- - 说是猫咪视角,结果变成无人机视角
- - 说是一镜到底,结果中间跳切
- - 室内路线穿过桌子、椅子、墙和人腿
- - 多人物互动里,角色脸和衣服互相漂移
- 原因通常不是缺少风格词,而是缺少行动轨迹。
- FPV 视频要写清楚的不是一张图,而是一段运动:谁在看,怎么走,经过谁,在哪里停,什么东西不能变。
- ---
- ## 适合什么场景
- - 咖啡厅、客厅、展厅、庭院、宫殿等室内一镜到底
- - 3 到 8 个角色依次互动的短视频
- - 猫、狗、机器人吸尘器、无人机、鸟、幽灵、车辆等非人类 POV
- - GPT Image / GPT-Image-2 首帧和人物参考图资产包
- - Seedance、Kling、Runway、Veo 等 image-to-video 工作流
- - 红线路径控制、世界地图飞行、城市到地标、峡谷穿越、赛车线路
- - 想把提示词从“视觉描述”变成“镜头调度”的创作者
- ---
- ## 工作原理
- 这个 Skill 会把一个 FPV 视频拆成 8 个问题。
- ```text
- 1. 摄像机是谁
- 2. 从哪里开始
- 3. exactly 有多少个主要人物或目标
- 4. 按什么顺序经过它们
- 5. 每一段路线是否物理可达
- 6. 每个停靠点发生什么互动
- 7. 哪些身份、服装、位置必须保持一致
- 8. 哪些东西绝对不能出现在最终画面里
- ```
- 然后它会选择两种路线模式之一。
- ### 编号停靠点
- 适合近距离人物互动、室内空间、咖啡厅、客厅、派对、展厅。
- 这类场景里不要默认画红线。红线很容易穿过桌椅、墙面和人腿,也容易残留在视频里。
- 更稳定的做法是,在首帧里放小编号 1、2、3,把角色顺序标清楚。真正的移动路线交给视频 prompt 约束。
- ### 红线路径控制
- 适合大世界路线、航拍地图、城市飞行、峡谷穿越、赛车线路、Seedance 2.0 path control。
- 这类场景的核心是路线几何。红线可以作为路径控制,但最终视频里必须完全消失。
- ---
- ## GPT Image 资产包
- 如果场景里有 N 个主要人物,Skill 默认生成完整资产包。
- ```text
- 1 张带编号首帧图
- N 张人物独立参考图
- 1 张可选干净首帧图
- ```
- 以 3 人咖啡厅为例:
- ```text
- 图片 1:咖啡厅猫咪视角首帧,带编号 1、2、3
- 图片 2:靠窗女生参考图
- 图片 3:咖啡师参考图
- 图片 4:角落男生参考图
- 图片 5:干净版首帧,去掉编号和所有标记
- ```
- 这样做的目的很简单:首帧管空间,人物参考管身份,干净首帧管最终输入。
- 如果只给一张图,视频模型很容易把路线、角色和画面标记混在一起。
- ---
- ## 安装
- ### 方式一:安装到 Hermes
- ```bash
- git clone https://github.com/zhouwei713/fpv-immersive-video-prompting.git \
- ~/.hermes/skills/creative/fpv-immersive-video-prompting
- ```
- 重启 Hermes,或开启新会话后即可使用:
- ```text
- fpv-immersive-video-prompting
- ```
- ### 方式二:作为通用 Skill 使用
- 如果你不使用 Hermes,也可以直接把 `SKILL.md` 放进 Claude、Codex、Cursor、OpenCode 或其他支持 Skill / long prompt 的 Agent 环境里。
- ### 方式三:只当提示词方法论参考
- 直接阅读:
- ```text
- SKILL.md
- skill/references/gpt-image-asset-packs.md
- skill/references/session-patterns.md
- ```
- ---
- ## 使用方式
- 你可以这样说:
- ```text
- 帮我做一个咖啡厅里 3 个人依次互动的 FPV 视频提示词,15 秒,猫咪视角
- ```
- ```text
- 做一个宫殿庭院 FPV,一镜到底经过 5 个角色,最后到池塘边
- ```
- ```text
- 我有一张世界地图,想让镜头沿红线从雪原飞到火山,帮我写 Seedance 2.0 prompt
- ```
- ```text
- 给我一套 GPT Image 生图 prompt,要包含首帧图、人物参考图和干净版首帧
- ```
- ---
- ## 仓库结构
- ```text
- .
- ├── README.md
- ├── README_EN.md
- ├── SKILL.md
- ├── skill/
- │ ├── SKILL.md
- │ └── references/
- │ ├── gpt-image-asset-packs.md
- │ ├── liyue-ai-redline-fpv-case.md
- │ ├── mayz-seedance-world-route-case.md
- │ ├── public-article-angle.md
- │ └── session-patterns.md
- ├── examples/
- │ ├── numbered-stop-example.md
- │ └── redline-route-example.md
- └── LICENSE
- ```
- ---
- ## 设计理念
- FPV 提示词不是把一张图说得更美。
- 它更像在设计一个小关卡:入口在哪里,路线怎么走,角色站在哪里,镜头能不能绕过去,最后有没有一个让观众理解空间的停顿。
- 只要运动没设计清楚,模型就会替你设计。它设计出来的结果,通常就是瞬移、穿墙、换脸和跳切。
- 这个 Skill 的价值,是把这套判断流程固定下来,让每次生成 FPV 视频前都先过一遍空间、路线、角色和物理限制。
- ---
- ## English
- For English documentation, see [README_EN.md](README_EN.md).
- ---
- ## License
- MIT License.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.