Cinematic Shot Architect

What it does

You give one scene idea, the clip lengths your video tool offers, a target length and the aspect ratio. The prompt returns four clips that open, build and end on your subject, each with one action, one camera move, a length from your list and a paste-ready prompt that works on its own.

A plain request usually gets you four pretty prompts. This one adds the parts that decide whether the clips survive the edit:

• A format line at the top of every paste-ready prompt: aspect ratio, clip length and "one continuous shot", so no clip comes out in the wrong shape or cuts inside itself.
• A STYLE LINE and a SUBJECTS block copied word for word into every clip, with enough detail for each recurring person and product that no clip has to rely on "she" or "it", because each clip is generated without seeing the others.
• Guards in every clip: the standard ones (no stray text or watermarks, natural hands, the product unchanged) plus the failure most likely in that particular shot, turned into a positive guard phrase.
• No invented logos or labels. A mark you didn't describe stays generic and goes on the list of things to confirm.
• Actions a video model can render, with a warning and a simpler version when a shot needs something delicate.
• The sum of the clip lengths checked against your target, which clip to generate first and what to check in it, and how to use a reference image so the product looks the same in every clip.

The whole answer is usually 600-1,000 words, most of it the paste-ready prompts.

Use it when

• You have an idea for a teaser, product clip or social video and want a shot plan before you spend generation credits.
• Your clips come out as one busy shot with too much happening, or the product or person changes between clips.
• You need several separately generated clips that look like one shoot.

Skip it when

• The scene is people talking. Use a dialogue prompt such as Veo Dialogue Scene, which sizes the lines to the clip.
• You already have a strong start image and only need movement. A motion-only image-to-video prompt is quicker.
• You need more than five shots or a full story. Plan the story first, then run this prompt once per scene.

The prompt

Plan a short video sequence that I will generate as separate clips in [VIDEO TOOL] and then cut together in this order. Each clip is generated on its own, without seeing the others, so every clip's prompt has to carry everything it needs.

Scene idea: [SCENE IDEA]
Clip lengths my tool offers: [CLIP LENGTHS, e.g. "5 or 10 seconds"]
Target total: [TARGET LENGTH, e.g. "about 30 seconds"]
Aspect ratio: [ASPECT RATIO]

If something is missing or unclear, make the simplest choice that invents nothing about my brand, use it, and list it under ASSUMED. Do not ask me questions first.

How to plan it:
1. Use 4 clips with an edit arc: clip 1 opens on the place and mood, clips 2 and 3 build, clip 4 ends on the subject or product. Use 3 or 5 clips only if no combination of 4 allowed lengths comes close to the target, and say why.
2. One visible action and at most one camera move per clip. If the idea needs more, split it or drop the weaker action. Prefer actions video models render well: slow, large and simple movements. If a clip needs a delicate one (fingers handling small parts, pouring, fast or crossing limbs, readable text), name the risk and offer a simpler version.
3. Use only the clip lengths I listed. Pick the lengths whose sum comes closest to the target and show the sum.
4. Write down what every clip has to share:
   - STYLE LINE: the look of the whole sequence: location, time of day, light source and direction, palette, lens feel.
   - SUBJECTS: one line per person, product or object that appears, recurring ones first, with the details a stranger would need to draw it. For a person: apparent age, hair, build, wardrobe and its colours. For a product: shape, size, colour, material, finish, and any marks exactly as I described them.
   Make these as long as they need to be. Their words go, unchanged, into every PASTE that shows that subject.
5. Logos, marks, labels and on-screen text: use only what I described. Never design a logo, slogan or label. If I mention a mark without describing its design, describe only what I did say (its colour, size or position), keep it small in frame, and put "logo design: confirm, or use a reference image" under ASSUMED. If I don't mention one, don't add one.
6. Describe only what a camera can see and a microphone can hear. Replace mood words with visible detail.
7. Guards. Every PASTE carries the standard guards: no on-screen text, captions, subtitles or watermarks (unless I asked for text); natural hands with five fingers whenever a hand is visible; the product's shape, colour and marks unchanged. Add the failure most likely in that particular shot (a small part morphing, colour drift, a face changing, a second camera move, a cut inside the clip) as a guard phrase that states the correct result positively, for example "the bottle stays matte black with one small white logo". Write the guards inside the PASTE, because some tools have no negative-prompt field.
8. Every format fact I gave (aspect ratio, clip length, and resolution or frame rate if I stated them) opens every PASTE, because each PASTE is used on its own.

Answer in this format, with no introduction. Usually 600-1,000 words, and the PASTE texts get the words first.
FORMAT: <aspect ratio> · <clip lengths used> · <resolution or frame rate, if I gave them> · <tool>
STYLE LINE: <the shared look>
SUBJECTS:
- <subject>: <identifying details>
CLIP <n> · <OPEN | BUILD | END> · <duration> s
SUBJECT: <subject + one action>
CAMERA: <lens feel + one move>
LIGHT: <source + direction>
MOTION: <what moves, most important first>
RISK: <the most likely failure in this shot> -> guard: "<guard phrase>"
PASTE: <one paragraph, usually 70-130 words, in this order: the format line ("<aspect ratio>, <duration>-second clip, one continuous shot, no cuts"); each subject named in full with its SUBJECTS details, never a pronoun that relies on another clip; the action; the setting and light; the camera move; the STYLE LINE; the standard guards and this shot's guard; one sound sentence if my tool generates audio (ambient sound and effects, and "no music" if music goes in the edit)>
(repeat for every clip)
TOTAL: <sum> s (target <target>)
TEST FIRST: <clip number, the riskiest one> - check <3 things> before generating the others
REFERENCE: <how to keep the subjects identical with my tool: one reference or start image used for every clip if the tool accepts one, or "text only: keep SUBJECTS identical"; say so if you are not sure what the tool accepts>
ASSUMED: <every choice I did not specify, each one to confirm, or "none">
EDIT: <a few lines, only if needed: what to add in the editor (text, logo, music, end card) and a pickup shot to generate if one clip fails>

How to use

1. Fill in the placeholders. [VIDEO TOOL] is the model you will generate with, for example "Kling VIDEO 3.0" or "Gemini Omni Flash in Google Flow". [CLIP LENGTHS] are the lengths that model offers (see How to run). If you know the resolution or frame rate you will use, add it to [ASPECT RATIO], for example "9:16, 1080p".
2. Paste the prompt into a chat model and read the plan once: four clips, one action each, and a TOTAL close to your target. Check that SUBJECTS describes your product and people the way they really look, and confirm or fix every ASSUMED line, logos first.
3. Set the aspect ratio, clip duration and resolution in the video tool's own settings too. The format line inside each PASTE backs them up; it does not replace them.
4. If the REFERENCE line says your tool takes reference images, prepare that image first: a photo of the real product, or one generated still you are happy with. Use the same image for every clip.
5. Generate the TEST FIRST clip. Check the three things it names. If one fails, fix SUBJECTS or the guard phrase now, while only one clip has been spent.
6. Generate the other clips by pasting each PASTE text.
7. Cut the clips in order and add text, logos and music in your editor, as the EDIT line suggests.

How to run (Oct 2026)

The chat model (writes the plan)

• Claude: paste it into a chat, or into a Project's instructions if you plan many scenes. Default effort is enough for four clips.
• ChatGPT: paste it into a chat or a Project. Instant or Medium on the Thinking slider is enough.
• Gemini: paste it into a chat, or save it as a Gemini app skill and call it with /.

The video model (makes the clips)

• Veo & Gemini Omni: Gemini Omni Flash (Gemini app, Google Flow, Gemini API) is Google's default. It works from text, an image or a reference, and has no negative-prompt field, so the guards inside PASTE are the right place for exclusions. Keep "one continuous shot, no cuts" in the format line so a clip does not cut inside itself. Veo 3.1 (Vertex AI, or inside Runway) makes 4, 6 or 8 s clips and takes up to 3 reference images; 1080p, 4K and reference images need 8 s clips. Both models generate audio with every clip, so keep the sound sentence in each PASTE ("Sound: kettle hiss and birdsong, no music"). Otherwise each clip gets different music, and you add one track in the edit instead.
• Kling VIDEO 3.0 (kling.ai, or inside Runway): clips of 3-15 s with native audio. Kling's guide orders a prompt as setting, characters, actions, then camera; if a clip ignores the camera move, put the setting sentence right after the format line. For a recurring person, bind an Element (2-4 reference images or a character video). Kling VIDEO 3.0 Omni is the version for reference-based generation and editing. When all four clips fit in 15 s, you can paste them into one generation with Custom Multi-Shot on, labelled "Shot 1:" to "Shot 4:".
• Runway Gen-4.5: 2-10 s at 720p, 24 or 25 fps. Paste PASTE into Text to Video. If you start from a still (Image to Video), keep only the camera and motion part, because Gen-4.5 image prompts describe motion only; the still carries the look.
• Seedance 2.0 / 2.5 (Jimeng AI, Doubao, BytePlus, or inside Runway): 2.0 makes up to 15 s and 2.5 up to 30 s in one run, and 2.5 takes up to 30 images as references. You can paste the clips as labelled shot blocks ("Shot 1: ...") in one generation; tag each reference (@Image 1) and say what it controls. Seedance's guide says exact timing is unstable, so let the model pace the shots.
• Aggregator apps (for example Higgsfield): choose the underlying model by name and follow its notes above.

Example

Input: [VIDEO TOOL] = Kling VIDEO 3.0 · [SCENE IDEA] = "A teaser for a hand-thrown ceramic mug with a sage-green glaze and a small stamped maker's mark; quiet early-morning kitchen; calm." · [CLIP LENGTHS] = 5 or 10 seconds · [TARGET LENGTH] = about 20 seconds · [ASPECT RATIO] = 9:16

Output (excerpt)

FORMAT: 9:16 vertical · 5 s clips · Kling VIDEO 3.0

STYLE LINE: Quiet kitchen at dawn, pale oak counter, low window light from the left, muted greens and creams, 50mm lens feel, shallow depth of field.

SUBJECTS:

• Mug: hand-thrown ceramic mug with a slightly uneven rim, matte sage-green glaze and one small stamped maker's mark near the base.
• Hand (clip 3): one adult hand, bare, short clean nails, cream knit sleeve.

CLIP 1 · OPEN · 5 s SUBJECT: steam rises from the mug on the counter CAMERA: 50mm feel, slow push-in LIGHT: low window light from the left MOTION: steam curls upward; camera creeps forward RISK: maker's mark multiplies or smears -> guard: "one small stamped mark near the base, unchanged" PASTE: 9:16 vertical, 5-second clip, one continuous shot, no cuts. A hand-thrown ceramic mug with a slightly uneven rim, matte sage-green glaze and one small stamped maker's mark near the base stands on a pale oak counter; steam curls slowly upward from it. Slow push-in toward the mug. Quiet kitchen at dawn, pale oak counter, low window light from the left, muted greens and creams, 50mm lens feel, shallow depth of field. No on-screen text, captions or watermarks; the mug keeps its shape, its sage-green glaze and one small stamped mark near the base, unchanged. Sound: soft kettle hiss and distant birdsong, no music.

CLIP 3 · BUILD · 5 s SUBJECT: the hand lifts the mug by its handle RISK: fingers merge around the handle -> guard: "one adult hand with five separate fingers, thumb resting on top of the handle" …

TOTAL: 5 + 5 + 5 + 5 = 20 s (target about 20 s) TEST FIRST: clip 3, the riskiest - check five fingers on the hand, a single maker's mark, the glaze still sage green. REFERENCE: photograph the real mug once in the same window light and use that photo as the reference for every clip; Kling VIDEO 3.0 Omni is the version for reference-based generation. ASSUMED: no faces in any clip; the hand in clip 3 is bare, with a cream knit sleeve (confirm). EDIT: add the shop name and an end card in the editor, not in the generation.

Illustrative example written by Meevra to show the target format; your model's wording will differ.

Troubleshooting

• The clips don't match. Make SUBJECTS more physical (material, exact colour, marks) and check that it is identical in every PASTE text. Use the same reference image for every clip if your tool takes one, and generate all clips with the same model and settings.
• A clip shows two actions or two camera moves. Cut SUBJECT to one verb and end the camera sentence with "no other camera movement".
• The logo or a small part morphs. Make the product smaller in frame, slow the camera, start from a reference photo, or leave the mark out of the generation and add it in the edit.
• It invented a logo, label or slogan. Reply: "Remove every logo, label or text I did not describe; describe the mark only by what I told you."
• A clip comes out in the wrong shape. Set the aspect ratio in the tool's settings as well as in the format line.
• Music changes from clip to clip. Your model generates audio. Add "no music" to every PASTE and lay one track in the edit.
• TOTAL is off target. Reply: "Recalculate using only [CLIP LENGTHS] and show the sum."
• The answer has an introduction or commentary. Reply: "Same plan, in the format only, no introduction."