Make an AI Short Film with a Grok Bot Team: 5 Bots, 6 Steps

Input a creative brief → 5 specialized Bots split into planning/directing/generating/editing → output an 18-second narrated short

Who

Creators who want multiple AI Bots to collaborate on a short film while controlling visual consistency and physical plausibility

You get

An 18-second, 5-shot rocket-launch short with male narration, produced in about 2 hours 45 minutes

You need

Grok Bot

Steps

1. Set up the team: create dedicated Bots and write role profiles

   Create 5 Bots in Grok Bot: Lead Bot (orchestrator), Planner, Director/QA, Generator, and Editor. Open the Bot info panel by clicking the Bot name at the top of the chat (or Cmd+Shift+I), then click the gear icon to set avatar, name, title, and description. The description is the ‘job description’ in four fixed blocks: name, duties, prohibitions, reporting. Source also gives the handoff file map between bots (brief.md / production-bible.md / renders/ / edit/ / log.md).

   Setup guidance · Source text。Grok Bot Director Bot description/setting field

   Copy prompt

   名：掌鏡師爺（导演 Bot）
   职责：
   • 专门做视频制作的“导演”：前期设计和质检。
   • 收到吹水王的策划和剧本后，出一份“制作圣经”：角色卡、场景卡、风格卡、分镜和镜头表（每镜几秒、景别、运镜、旁白对位）。
   • 质检：神筆馬良每出一个镜头，先检查物理和透视（没有穿插、没有浮空、单一光源阴影一致、比例正常）、角色一致性、镜头衔接。不合格就写具体修改意见打回去，每个镜头最多改两次。
   • 全程记录：每步的指令、结果、改动写进项目文件夹的记录文件，给吹水王写教学文。
   禁止：不发帖、不登录 X、不处理邮件；不模仿真人肖像、不侵犯版权。
   汇报：做完工作或有重要进度，同时向我和主 Bot“大掌柜”汇报（简短：做了什么、结果、有没有要我决定的事）。

   Setup guidance · Source text。Grok Bot Generator Bot description/setting field

   Copy prompt

   名：神筆馬良
   职责：专门制作视频和图片素材：长文封面、配图、短片、示范视频、旁白、配乐；由主 Bot 统筹派工，收到主题后先建议画面、脚本和规格，再出素材。
   禁止：不代为发布任何社交帖；不处理邮件、日程或 X 文案；不用真人肖像冒充真人，不侵犯版权。

   Execution prompt · Source text。团队交接约定 / handoff map between bots

   Copy prompt

   brief.md             吹水王 → 掌鏡師爺：策划、剧本、旁白、物理提醒
   production-bible.md  掌鏡師爺 → 神筆馬良：世界规则、角色卡、场景卡、风格卡、分镜镜头表、逐镜提示词、质检清单
   renders/             神筆馬良 → 掌鏡師爺：每镜的关键帧、视频，以及记录提示词原文的 prompts.md
   edit/                掌鏡師爺 → V仔：剪辑交接单 handoff-to-Vzai.md；V仔 → 掌鏡師爺：剪辑方案、成片、检查报告 check-v1 到 v4
   log.md               所有人：每一步的时间、结论、改动和原因

2. Planning and script (Planner Bot)

   The Planner Bot locks down ‘what, for whom, how long’ and outputs brief.md. Core fields: audience profile, platform and specs, core message, style, and physics reminders. Physics reminders must go into the plan; the most common AI-video giveaways are floating objects, wrong smoke direction, and inconsistent light sources.

   Execution prompt · Source text。brief.md content, passed to Director Bot as input

   Copy prompt

   观众：对 AI、SpaceX、Tesla 有兴趣，想学用 AI 自己做片，未必有剪辑经验
   平台：X；时长约 18 秒（上限 20 秒）；画幅 16:9
   核心信息：一条有电影感、物理合理的短片，可以由几个 Bot 分工、全部用 AI 做出来
   风格：写实电影感夜景，光源只有发射塔灯和之后的火焰；画面不放 logo、不放文字
   物理提醒：白汽往下沉，不要往上飘；起飞初段要慢；光弧由低到高、向一边倾斜

3. Director writes the production bible (Director Bot)

   The Director Bot receives brief.md and outputs production-bible.md, translating the script into machine-executable, human-checkable specs. Key actions: lock set geography and direction (all cameras south, tower on left, rocket flies right); give exact numbers (rocket 70m×8m, tower 50m, per-second ascent targets); redesign shots (e.g., shot 3 changed to 35mm medium shot); write consistency anchors (rocket/tower/pad/lighting/style in English paragraphs pasted per shot); set rules (no video until keyframe passes QA; video prompts describe motion only; max 2 revisions per shot, then redesign).

   Execution prompt · Source text。Execution instruction sent to Director Bot

   Copy prompt

   根据 brief.md 写制作圣经：世界规则、角色卡、场景卡、风格卡、分镜镜头表、逐镜生成提示词、逐镜质检清单。每镜最多改两次，第二次不过就改分镜。每张图、每条视频按清单逐项答“是/否”，任一项“否”就退回。

   Execution prompt · Source text。Style anchor, pasted verbatim into each keyframe prompt

   Copy prompt

   photorealistic cinematic night documentary footage, 16:9, fine film grain, deep teal-blue shadows with detail, realistic scale, sharp straight structural lines, real night photograph shot on a full-frame cinema camera at high ISO, natural grime and weathering on concrete and steel, not a 3D render

4. Master reference image (Generator + Runway)

   Generate one wide shot showing rocket, tower, and pad clearly to establish the film’s look. Use Runway image generation (nano-banana-pro, 16:9, 2K). Workflow: text-to-image first, then local edits; do not use reference images to fix composition (errors propagate). No keyframes or video until the master passes QA.

   Execution prompt · Source text。Master-phase instruction sent to Generator Bot

   Copy prompt

   第一步只出母版参考图（第 7.2 节），存到 renders/master/，并记录提示词和模型；母版质检过关前不做关键帧和视频。

5. Per-shot keyframes (Generator + Grok Imagine)

   Generate a static first frame for each shot; only animate after QA pass. For anchored keyframes always use Grok Imagine Quality 2.0 (Speed mode ignores long prompts). Output size 1792×1008. When uploading an image for local edit, the Quality 2.0 toggle disappears and only 1 image is returned. Edit pattern: ‘first say nothing may change, then say only these two things.’ For close-ups, describe only visible elements; do not paste the full anchor set.

   Execution prompt · Source text。Grok Imagine local-edit input, used for shot 1 final

   Copy prompt

   Edit this image. Keep everything exactly the same: composition, camera, lighting, tower, truss arm, floodlight masts, launch table, trench openings, puddles, reflections and horizon. Change only two things: 1) the engine nozzles are recessed inside the launch table and not visible; the rocket's white base meets the table top at a clean flush seam with a dark contact shadow. 2) make the stars sparse and faint.

   Execution prompt · Source text。Grok Imagine local-edit input, used for shot 4 redesign

   Copy prompt

   Keep everything in this image exactly the same, including the camera, framing, tower, launch table, floodlight masts, steam clouds, ground and sky, and keep the rocket's size and shape exactly the same. Change only two things. First, move the whole rocket straight down so its white base sits directly on the top of the launch table, with only a thin bright orange glowing seam where the base meets the table; the flame is hidden inside the table. Second, the tower's left side, facing away from the rocket, is in cool blue-gray shadow, and only the tower's right side facing the rocket is lit warm orange.

6. Image-to-video (Generator + Grok Imagine)

   Animate approved first frames into video. Generate 2 clips per shot and pick 1. Settings: Grok Imagine video mode, 6 seconds, 720p, ambient sound included (delete in edit). Video prompts describe motion and boundaries only (e.g., dolly percentage, ripple intensity), do not repeat the image description. QA: compare first/last frames for motion amplitude, measure per-second ascent, frame-by-frame brightness check for blowouts, watch every clip in full. After all 5 shots are approved, the Director writes the edit handoff edit/handoff-to-Vzai.md and passes to Editor.

   Execution prompt · Source text。Grok Imagine video motion prompt, shot 1

   Copy prompt

   Steady slow dolly forward, clearly moving closer to the launch pad, about five percent closer by the end. The puddle surface stays almost mirror-still with only faint slow wind ripples; reflections stay aligned beneath their lights. The air is completely clear and still. The rocket, tower, arm and lights stay perfectly still and unchanged. Calm, quiet night.

   Execution prompt · Source text。Grok Imagine video motion prompt, shot 2 final version

   Copy prompt

   Locked-off camera, almost static, with only a very slight slow tilt down of about five percent. The framing always stays on the middle section of the rocket; the bottom of the rocket is never shown. The heavy cold vapor keeps sliding straight down the rocket surface and flows out of the bottom edge of the frame. The rocket, the arm and the tower do not move. The air around is still.

   Execution prompt · Source text。Grok Imagine video motion prompt, shot 2 test version (failed, for comparison)

   Copy prompt

   Very slow camera tilt down by a few meters, following the vapor. The heavy cold vapor keeps sliding down the rocket surface and falls out of the bottom of the frame. The rocket, the arm and the tower do not move. The air around is still.

7. Edit, voice-over, and mix (Editor Bot)

   Cut the 5 clips into an 18-second film per the Director’s handoff sheet. The sheet locks: cut list (files, in/out points), hard cuts only, no speed change, no color grade, no subtitles, narration script, sound design. Editor submits a cut plan first; Director approves before cutting begins. Frame sync: statistically analyze shot 3 brightness and orange pixels to find the true ignition frame (source frame 41), place it at timeline frame 240 = 10.000 s. Narration: edge-tts Mandarin male voice YunjianNeural; the word ‘火’ starts at exactly 10.000 s. Sound and music mixed locally with ffmpeg; music ducks ~7 dB under narration. Per-version checks with ffprobe for specs, cut-point frames, loudness; 3 speech-recognition models transcribe the output for review instead of listening. Target: -14.0 LUFS, true peak -1.4 dBTP.

No prompt needed — just follow the steps.