# How to Write Prompts for AI Video

A prompt method for AI video that covers subjects, camera and motion, image inputs, consistency, audio, iteration cost and rights, with how Runway, Flow, Kling, Luma and Pika read prompts differently.

An AI video prompt works when it reads like a shot description a camera operator could follow. Runway, Google, Kling AI and Luma each publish prompt rules that point the same way, and Pika routes a prompt to the model you choose.

- Write an AI video prompt as one scene with a visible subject, one action, setting, camera move, light and style.
- Write motion, not appearance, when you start from an image.
- Phrase everything positively, because Runway says negative phrasing is unsupported.
- Lock consistency with reference images and the same details repeated in every prompt.
- Change one element per take and budget takes against your credits.

This applies to generative scene tools that take text or image prompts and return short clips, such as Runway, Google Flow, Kling AI, Luma and Pika. Runway's Gen-4 makes 5- or 10-second clips, and Kling AI advises visuals that complete within 5 to 10 seconds. Each model reads prompts differently, so a prompt tuned for one tool needs adjusting for another. Avatar presenter tools follow scripts rather than scene prompts, which is why [AI avatar video tools](https://growthtribune.com/best-ai-avatar-video-tools/) get their own comparison.

## What Should An AI Video Prompt Include?

Google, Kling AI and Luma document the same five building blocks for a video prompt (who or what is on screen and what it does, where it happens, how the camera sees it, the light and style, and, on models that generate sound, the audio), and Kling's official formula makes only the subject, its movement and the scene mandatory.

If you are new to [AI video generation tools](https://growthtribune.com/content-audience/ai-video/), start with [what AI video generation is](https://growthtribune.com/what-is-ai-video-generation/) and come back for the prompt itself. The table turns the shared anatomy into something you can write against.

**The five building blocks of an AI video prompt**

| Element | What to write | Example |
| --- | --- | --- |
| Subject and action | Who or what, doing one thing | A barista pours latte art |
| Scene | Where it happens, with detail | Sunlit café counter, morning |
| Camera | Shot size and one move | Close-up, slow push-in |
| Lighting and style | Light source and visual look | Warm window light, cinematic |
| Audio | Ambience, effects or dialogue | Espresso hiss, quiet chatter |

The vendors agree on the elements but not the order. Kling AI's formula starts with the subject and its description, adds the subject's movement and the scene, and treats camera language, lighting and atmosphere as optional. Luma's Dream Machine guide runs subject, action, subject details, scene and style, then ends with a camera move and a reinforcer. Google's Flow tips list subject and action, composition and camera, location and lighting, style and audio.

The practical rule is to write the mandatory three first. A prompt with a clear subject, one movement and a described scene gives every model enough to build a shot. Camera, light and style then refine it. Audio only matters on models that generate sound, such as Veo 3 in Google Flow and Kling VIDEO 3.0.

Kling AI also asks for simple words and visuals that complete within 5 to 10 seconds. That limit shapes everything else in the prompt. One subject doing one thing in one place fits a clip of that length, and a sequence of events does not.

## How Do You Describe Subjects In An AI Video Prompt?

A subject in an AI video prompt should be described by what a camera could record, so Luma's guide asks for visible detail such as shaking hands and tense shoulders instead of a word like nervous, and Kling AI warns that exact object counts may not stay consistent. Describe appearance and behaviour, never inner states.

Four habits keep subjects readable:

- **Describe appearance concretely.** Kling AI's formula pairs the subject with a description in simple words, such as "a woman in a navy blazer with short grey hair".
- **Show behaviour, not emotion.** Luma asks for what can be seen. A feeling gives the model nothing to animate, while a gesture does.
- **Give the subject one main action.** Runway's Gen-4 makes 5- or 10-second clips and treats each generation as a single scene. One action fits that window.
- **Use position for several subjects.** Runway suggests positional language such as "The subject on the left walks forward", so the model knows who does what.

Avoid promising exact numbers. Kling AI notes that the model handles numbers poorly, so a prompt for five people around a table may not hold that count. Write "a small team around a table" when the count does not matter, and start from an input image when it does, since the image sets the subjects.

> **Tip:** Rewrite feelings as actions a camera can see. Before, "The founder feels nervous before the pitch." After, "The founder's hands shake as she straightens her notes, her shoulders tense." The second version gives the model movement to animate as well as the mood you want.

## How Do You Write Camera Movement In An AI Video Prompt?

Runway's Gen-4 guide asks for the camera's movement style (locked, handheld, dolly or pan), whether it tracks the subject, and any focus shift, and tells writers to phrase stillness positively as "Locked camera. The camera remains still." rather than saying what the camera should not do. Name one move per clip.

Camera direction gets clearer once you separate three things that prompts often blur together:

- **Shot size.** How much of the scene the frame holds. Google's Flow tips recommend terms such as "wide shot" and "close-up".
- **Camera movement.** How the camera travels. Runway, Kling AI and Luma each list movement terms, from a locked camera to a tracking shot.
- **Subject and scene motion.** What moves inside the frame, which Runway treats as separate from camera motion.

**Camera terms named in vendor prompt guides**

| Term | What it does | Listed by |
| --- | --- | --- |
| Wide shot | Shows the full setting | Google Flow |
| Close-up | Fills the frame with detail | Google Flow |
| Locked camera | Camera stays still | Runway |
| Handheld | Loose, human-held movement | Runway |
| Dolly | Camera travels on a path | Runway |
| Pan | Camera turns sideways | Runway, Kling AI |
| Tilt | Camera turns up or down | Kling AI |
| Slow push-in | Camera moves toward subject | Kling AI |
| Tracking shot | Camera follows the subject | Google Flow, Kling AI, Luma |
| Crane down | Camera lowers vertically | Luma |
| Camera circles slowly | Camera orbits the subject | Luma |
| Aerial view | Looks down from above | Google Flow |
| Focus shift | Focus moves within frame | Runway |

Kling AI's guide frames camera direction around what the viewer will see, naming moves such as a slow push-in, pan, tilt or tracking shot. Luma's examples include crane down, tracking shot and "camera circles slowly". The wording differs, but each guide uses plain film language rather than technical settings.

Runway adds a subtler point about motion. Insinuated motion, such as "The subject runs across the dusty desert", often looks more natural than a description of the environment moving. Let the subject's action imply the motion of the scene.

A useful decision rule follows from Runway's advice to treat each 5- or 10-second generation as a single scene. Give each clip one camera move. A push-in and a pan in the same clip compete for the same few seconds. When the shot needs both, generate two clips and cut between them.

## What Is A Good AI Video Prompt Template?

A reusable AI video prompt template follows the order Luma documents for Dream Machine (main subject, action, subject details, scene, style, camera move and a reinforcer), which keeps the camera instruction last where it is easy to swap between takes. Fill each slot in plain words and leave out any slot the shot does not need.

**AI video prompt template**

```
TEXT-TO-VIDEO
Main subject, action, subject details, scene, style, camera move, reinforcer

EXAMPLE
A matte-black wireless speaker rotates slowly on a round pedestal, fine fabric texture and a small glowing status light, in a bright minimalist studio with a white backdrop, clean commercial product style, camera circles slowly, crisp and evenly lit.

IMAGE-TO-VIDEO (the image sets the look)
The subject rotates slowly on its pedestal. The camera circles slowly around it.
```

What each slot does:

- **Main subject and action.** One subject doing one thing, two of the three parts Kling AI's formula makes mandatory.
- **Subject details.** Visible features a camera could record.
- **Scene and style.** Where it happens and the look. Google's Flow tips include non-realistic styles such as stop motion or clay animation.
- **Camera move.** One move from the camera table above.
- **Reinforcer.** Luma's name for the closing slot. Use it to restate the quality the shot must keep.

The image-to-video version drops every visual detail and keeps only motion. For a background clip on a landing page, Luma notes that adding the word "loop" makes a looping clip.

## How Do You Write A Prompt For Image To Video?

When a video starts from an image, Runway's guide says the prompt should describe motion only, because the input image already sets the subject, composition, colour, lighting and style, and restating those visible details can reduce how much the clip moves. The image does the describing, and the prompt does the directing.

An image-to-video prompt is therefore shorter than a text prompt. Runway suggests referring to subjects generally, as "the subject", and using positional language when several subjects appear. A text prompt for a product shot might run to a full paragraph, while the image-to-video version of the same shot can be two sentences about what moves and how the camera follows.

The input image matters as much as the words. Runway asks for a high-quality input image free of artifacts.

Several tools take more than one frame. Luma lets start and end keyframes guide a generation, so the clip travels between two compositions you choose. Google Flow's Frames to Video starts a clip from an uploaded image or from a frame of a previous generation, which lets one clip pick up where the last ended. Pika lists Pikaframes among its own video models, alongside Pika 2.5.

Luma's Ray3 also accepts direction that is not text at all. Visual Annotation lets you scribble directions on the start frame to guide motion and camera, which helps when a path is easier to draw than to describe.

## How Do You Keep Characters Consistent In AI Video?

Consistency across AI video shots comes from reference images plus repeated text. Runway accepts up to three saved references called with an @ tag, Google Flow up to three ingredients per prompt, and both Google and Kling AI tell writers to repeat every essential detail, such as clothing, hairstyle and props, in each new prompt.

Each tool offers its own reference feature:

- **Runway References.** Up to three references per generation. Saved references are tagged and called in the prompt with @, for example @bryan.
- **Google Flow ingredients.** An ingredient is a consistent character, object or style reference, with up to three per prompt. When Google published its Flow tips, Ingredients worked only with Veo 2.
- **Kling AI Element Reference.** Kling AI's guide recommends it for recognisable subjects, alongside identical clothing, hairstyle, props and role in every shot.
- **Pika Character Studio.** An app for building reusable characters.
- **Luma Draft Mode and mastering.** A chosen draft can be mastered into a 4K HDR version that keeps the same identities, motion and composition.

The same discipline applies to products. A product ad that cuts between three clips needs the same colour, material and label described in identical words each time, plus the same reference image. Changing "matte black" to "dark grey" between prompts invites the model to change the product.

Teams producing brand-consistent client work at volume should weigh these features when choosing a platform, a trade-off our comparison of [AI video tools for agencies](https://growthtribune.com/best-ai-video-tools-for-agencies/) covers.

> **Tip:** For Runway References, use evenly lit source images with a neutral expression. Describe shoes or trousers when you want a full-body shot, reuse a good output as a new reference, and tag every upload you want to keep, because untagged images disappear on refresh.

## How Do You Prompt AI Video For Dialogue And Sound?

Dialogue prompts work when the line fits the clip, so Google's Flow tips warn that Veo 3 is less likely to produce speech if the requested dialogue does not fit in the 8-second clip, and Kling AI asks for the speaker's name, line and delivery to sit together in the prompt. Keep lines short and attributed.

In Google Flow, selecting Veo 3 generates audio, which Google describes as experimental. You can prompt ambient noise, sound effects or dialogue, and specify tone, emotion or accent.

Kling VIDEO 3.0 Native Audio supports dialogue, effects and ambience in Chinese, English, Japanese, Korean and Spanish. Its guide keeps each speaker's name, line and delivery together, which keeps who says what unambiguous.

A sound-ready prompt adds two short lines to the scene:

- **Ambience.** Quiet open-plan office, keyboards tapping, a phone ringing in the distance.
- **Dialogue.** The manager on the left says warmly, "We shipped it."

The dialogue line names the speaker by position, gives the delivery and keeps the words short enough to fit the clip. Pika lists Pika Soundtrack and SFX Studio among its apps for adding sound.

## How Does Prompting Differ Between AI Video Tools?

The five generators covered here read the same prompt differently. Runway wants motion first and does not support negative phrasing, Kling AI's formula starts from a described subject, Luma ends with a camera move, Google Flow adds audio through Veo 3, and Pika routes the prompt to whichever of its in-house or third-party models you pick.

**How five AI video tools read a prompt**

| Tool | Prompt emphasis | Image or frame input | Consistency feature |
| --- | --- | --- | --- |
| Runway | Motion first, positive phrasing | Image sets the look | References with @ tags |
| Google Flow | Five parts, including audio | Frames to Video | Up to three ingredients |
| Kling AI | Subject first, camera optional | Element Reference | Identical details per shot |
| Luma Dream Machine | Ordered, camera move last | Start and end keyframes | Draft, then master |
| Pika | Follows the chosen model | Pikaframes model | Character Studio |

**Runway.** Runway's Gen-4 guide starts from a simple prompt that captures only the core motion, then adds subject motion, camera motion, scene motion and style one at a time. From an image, the prompt describes motion only. Gen-4 References keep a character or product consistent with up to three tagged references. Gen-4 makes 5- or 10-second clips, so each generation is one scene. Runway meters generation in credits, and its pricing page converts them to seconds, with 625 credits buying 52 seconds of Gen-4.5 or 104 seconds of Gen-4 Turbo. Every discarded take spends footage you could have kept.

**Runway pricing**

| Plan | Price | Billing | Notes |
| --- | --- | --- | --- |
| Free (Runway Creative) | $0 | per month |  |
| Standard (Runway Creative) | $15 | per month |  |
| Pro (Runway Creative) | $35 | per month |  |
| Max (Runway Creative) | $95 | per month |  |
| Enterprise (Runway Creative) | Custom pricing | per custom |  |

**Google Flow.** Google's Flow tips ask for subject and action, composition and camera, location and lighting, style and audio, and suggest Gemini for refining prompts. Frames to Video starts a clip from an uploaded image or a frame of an earlier generation, and ingredients hold up to three characters, objects or styles steady. Selecting Veo 3 adds experimental audio, including dialogue, in an 8-second clip. Extend continues action from a clip's final frames, though Google noted that Ingredients and Extend then worked only with Veo 2. What Flow costs depends on the plan in its pricing table.

**Kling AI.** Kling AI's official formula is subject with description, subject movement and scene with description, with camera language, lighting and atmosphere optional. Element Reference keeps recognisable subjects steady, and Multi-Shot prompts set the scene first, then list shots in order. Kling VIDEO 3.0 Native Audio adds dialogue, effects and ambience in five languages. Kling AI advises visuals that complete within 5 to 10 seconds. What Kling AI costs depends on the plan in its pricing table.

**Luma Dream Machine.** Luma's guide orders the prompt as subject, action, subject details, scene, style, camera move and reinforcer. Start and end keyframes guide a generation, and Visual Annotation lets you draw on the start frame. Draft Mode explores variations up to 5x faster, and mastering a chosen draft into 4K HDR keeps the same identities, motion and composition. Extend can take a 5-second clip up to 30 seconds, though quality may decline the longer you extend. Luma meters Dream Machine in credits and lists the credits each generation type uses. The Plus plan includes 10,000 credits and lists commercial use.

**Luma Dream Machine pricing**

| Plan | Price | Billing | Notes |
| --- | --- | --- | --- |
| Plus | $30 | per month |  |
| Pro | $90 | per month |  |
| Ultra | $300 | per month |  |
| Enterprise | Custom pricing | per custom |  |
| Team | Custom pricing | per custom |  |

**Pika.** Pika presents itself as a multi-model platform, so a Pika prompt follows the conventions of the model you pick, whether one of Pika's own models or a third-party model such as Seedance, MiniMax, Wan or Gemini Omni. Write the shared anatomy first, then adjust it to the chosen model. Pika offers its Pikaframes model and a Motion Transfer app for work that starts from existing material, Character Studio builds reusable characters, and Pika Soundtrack and SFX Studio add sound. Commercial use is included on the Creator and Fancy plans, and the plans below them grant no commercial rights.

**Pika pricing**

| Plan | Price | Billing | Notes |
| --- | --- | --- | --- |
| Free | $0 | per month | per credit |
| Starter (billed annually) | $8 | per month, billed annually | per credit |
| Starter | $10 | per month | per credit |
| Creator (billed annually) | $28 | per month, billed annually | per credit |
| Creator | $35 | per month | per credit |
| Fancy (billed annually) | $76 | per month, billed annually | per credit |
| Fancy | $95 | per month | per credit |
| Enterprise | Custom pricing | per custom | per contact |

Readers still choosing between these tools should start with our comparison of [AI video generators for marketing teams](https://growthtribune.com/best-ai-video-generators/).

## What Does An AI Video Workflow Look Like?

With [63% of video marketers having used AI video tools to create or edit marketing videos, up 12 points in a year](https://www.wyzowl.com/sovm-results-2016), according to Wyzowl's 2026 survey of 266 respondents, a repeatable prompting loop matters more than any single prompt. Draft cheaply, change one element per take, and master only the shot worth keeping.

- [ ] Define the one shot and the job it does in the edit
- [ ] Lock references for every character and product in the series
- [ ] Write the base prompt with subject, action and scene only
- [ ] Generate cheap drafts first, for example in Luma's Draft Mode
- [ ] Change one element per take, as Runway's guide advises
- [ ] Set an iteration budget in credits before the first take
- [ ] Extend or chain clips from final frames or keyframes
- [ ] Master the chosen take at full quality
- [ ] Edit the sequence and add any disclosure the platform requires

Budget against the credit meter. On Runway, 625 credits buy 52 seconds of Gen-4.5, so four tries at a 5-second shot use 20 of those 52 seconds. Luma's Draft Mode, which generates up to 5x faster, exists for exactly this exploration step, and mastering keeps the same identities, motion and composition once a draft earns it.

Three measures show whether the loop is working:

- **Seconds kept per credits spent.** The share of generated footage that reaches the final cut.
- **Iterations to an approved take.** A falling number means your base prompts are getting sharper.
- **Consistency across a series.** Whether characters and products still match from the first clip to the last.

The same Wyzowl survey found that 89% of consumers say video quality affects how much they trust a brand. That is why the master step deserves its own line in the budget.

## What Mistakes Should You Avoid In AI Video Prompts?

One prompt failure Runway calls out directly is asking the model what not to do. Runway's Gen-4 guide states that negative phrasing is not supported and can produce unpredictable or opposite results, so "no camera movement" should be written as "Locked camera." Most other failures come from asking one clip to do too much.

Each mistake has a direct fix:

- **Negative phrasing.** Write the positive version of what you want to see.
- **Several scene changes in one clip.** Runway advises against multiple scene changes in one generation. Split them into separate clips.
- **Abstract or emotional words.** Runway advises against abstract concepts, and Luma asks for what can be seen. Swap feelings for visible actions.
- **Command-style prompts.** Runway advises against conversational or command-style prompts. Instead of "add a dog", describe the scene with the dog already in it.
- **Re-describing the input image.** Runway says restating visible features can reduce motion. Describe motion only.
- **Exact counts and complex physics.** Kling AI notes that the model handles numbers poorly and that complex physical motion is hard to generate. Keep counts loose and actions simple.
- **Extending too far.** Luma says quality may decline the longer you extend a clip. Extend only to the length the edit needs.

> **Warning:** Do not rescue a bad take by stacking more instructions onto the same prompt. Runway's guide recommends starting with the core motion and adding one element at a time, so you can see what each addition changes. Strip the prompt back and rebuild it step by step.

## Who Owns AI Generated Video?

Under the [U.S. Copyright Office's January 2025 report, prompts alone do not give enough human control](https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf) to make the prompter the author of an AI video, so the protectable parts of a marketing clip are the human edits, arrangement and original inputs around the generation. Commercial use is a separate question, set by each tool's plan.

The report, Copyright and Artificial Intelligence Part 2, ties its conclusion to current generally available technology. It says creative selection, arrangement or modification of outputs, and a person's own expressive inputs, can be protected. For a marketing team, protection sits in the cut, the sequence and the original footage, product shots or artwork you feed in.

Commercial rights depend on the plan, not the prompt. Pika includes commercial use on its Creator and Fancy plans and grants no commercial rights on the plans below them. Luma lists commercial use on its Plus plan. Read the plan terms of the tool you use before an ad goes live.

Rights to anything you upload are yours to secure. An image of a real employee, customer or spokesperson used as a start frame or reference should be cleared with that person for the use you intend.

**Disclosure on YouTube.** [YouTube requires creators to disclose realistic altered or synthetic content](https://blog.youtube/news-and-events/disclosing-ai-generated-content) that a viewer could mistake for a real person, place, scene or event, such as a realistic likeness or a generated voice. Clearly unrealistic content such as animation, minor colour or lighting edits, and AI used for productivity do not need disclosure.

> **Note:** YouTube shows the label in the video description, and more prominently on sensitive topics. A product clip in a clay animation style would not need a label, while a realistic generated spokesperson would.

This section is general information, not legal advice.

## FAQ

**How long should an AI video prompt be?**

An AI video prompt should describe one scene that fits the clip, because Runway's Gen-4 makes 5- or 10-second clips and Kling AI advises simple words and visuals that complete within 5 to 10 seconds. Start with subject, action and scene, then add camera, light and style one element at a time.

**Can you use negative prompts in AI video?**

Write the positive version instead. Runway's Gen-4 guide states that negative phrasing is not supported and can produce unpredictable or opposite results. To stop the camera moving, write "Locked camera. The camera remains still." To keep a background empty, describe the empty background you want rather than listing what to leave out.

**Can you write AI video prompts with ChatGPT or Gemini?**

Yes, as a drafting aid. Google's own Flow tips recommend using Gemini to refine video prompts. Whichever assistant drafts the prompt, edit the result against the target tool's rules, such as motion only for image-to-video in Runway or Kling AI's subject, movement and scene formula, before spending credits on a take.

**Can one prompt make several shots?**

In Kling AI, yes. Its Multi-Shot prompts define the setting first and then list each shot in order with its shot size. Runway takes the opposite approach and treats each 5- or 10-second generation as a single scene, so in Runway you write one prompt per shot and join the clips in the edit.

## Sources

- [Video Marketing Statistics 2026 (12 Years of Data) | Wyzowl](https://www.wyzowl.com/sovm-results-2016), wyzowl.com
- [Copyright and Artificial Intelligence, Part 2: Copyrightability](https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf), copyright.gov
- [How we're helping creators disclose altered or synthetic content](https://blog.youtube/news-and-events/disclosing-ai-generated-content), blog.youtube

---
Source: https://growthtribune.com/how-to-write-prompts-for-ai-video/