# Kling AI Prompt Guide: Kling's Own Rules, Formulas and Negatives

> How to prompt Kling AI with Kling's own published rules: the subject, movement and scene formula, Kling 4.0 asset roles, two negatives and camera tips.

_Source: https://professionalstoolkit.com/articles/kling-ai-prompt-guide — The Professional's Toolkit · updated November 2026_

---


**50-100** — Words Kling recommends for most scene prompts; past 150 words tends to give poorer results.



- **2** — Negatives Kling's 4.0 guide calls necessary
- **1** — Main camera move per shot, per Kling
- **8,000** — Token limit for a Kling 4.0 prompt


## The formula Kling publishes, and which one to use

Kling's own answer is short. Its text-to-video guide builds a prompt from three parts: the subject, what the subject does, and the scene, with camera language, lighting and atmosphere added when they matter [1]. For Kling 4.0, the newest model since its release to all users on October 15, 2026, the advice shifts slightly: open with one sentence that names the element, scene, action and emotion, then add only details that change the result. Keep most prompts between 50 and 100 words. Kling's advanced-techniques blog says results tend to get worse past 150. And only two negatives are called necessary: no subtitles and no background music. Kling has published its formula in several places, and the versions differ in length more than in substance.

| Kling source | Parts of the prompt |
|---|---|
| Text-to-video guide (same in the VIDEO O1 guide) | Subject + movement + scene; optional camera language, lighting, atmosphere |
| Image-to-video guide | Subject + movement, then background + movement; the image supplies the scene |
| AI video prompt blog | Subject, action and motion, environment, camera movement, lighting, style and pacing, constraints |
| Camera control blog | Subject, environment, action, camera movement, visual style |
| Horror prompts blog | Subject, action, setting, style, audio; three to five core elements |
| Kling 4.0 user guide | One sentence (element, scene, action, emotion), then camera, timing, audio, preservation |

The common core is a subject doing one thing in a place. Everything else is optional, and Kling's guidance leans toward adding it sparingly. The text-to-video guide advises keeping visuals simple enough to complete within 5 to 10 seconds, a rule written for the shorter clips of older models; Kling 4.0 now runs 3 to 30 seconds per generation and, with multi-shot on by default, splits a long take into shots by itself. The same guide warns that video models are not sensitive to numbers. Ask for ten puppies, and the count will drift. Complex physics, such as a bouncing ball, is hard as well.

## Writing for Kling 4.0: one sentence, then roles

Kling 4.0 accepts prompts of up to 8,000 tokens, but Kling does not treat that limit as a target. Its 4.0 user guide asks for one opening sentence that carries element, scene, action and emotion, followed only by details that matter: camera work, timing, audio, and what must be preserved [2]. The guide also asks you to replace abstract wording with concrete description. A request for a premium look or a cinematic feel tells the model little; the lighting, the materials and the composition tell it a lot. Kling's own product-shot example follows that pattern: a glass perfume bottle turning slowly on wet black stone, a macro push-in, the label kept unchanged.

### Name every asset's job

In Omni Reference, the default entry point, you can attach up to 15 assets: up to 10 images, 5 videos totalling no more than 30 seconds, 7 elements and voice references. Kling numbers them in upload order, as @Image1, @Image2 or @Video1, and asks you to state each one's role in the prompt. Skip the roles and the model guesses how to use each asset. Results then get less predictable. An editing prompt needs the same discipline: say what changes and what stays. Without both halves, Kling warns, a reference-style prompt may regenerate the clip instead of editing it.


> 💡 **Give every reference a role:** Kling numbers assets in upload order (@Image1, @Video1); say what each one is for, or the model guesses and results get less predictable.


Keyframes, the second entry point, works from a timeline instead. Kling's sample maps a description of each keyframe to a time range across a 15-second clip, which suits teams that already have a storyboard. You can place up to 10 keyframe images, including first, last and middle frames, but Keyframes and Omni Reference cannot be combined in one task. Kling's rule of thumb is Keyframes when visuals must be locked at specific moments, and Omni Reference when you combine references or explore ideas. Kling 4.0 Flash, the cheaper variant capped at 720p, offers neither start and last frames nor multiple keyframes.

## Negative prompts: two that matter, one exception

People searching for a Kling negative prompt usually expect a long exclusion list. Kling's 4.0 guide says the opposite: only two negative instructions are truly necessary, **no subtitles and no background music**. Both cover things Kling 4.0 can produce on its own, since it generates an audio track and renders on-screen text. The guide adds one exception for style. When you ask for animation, explicitly exclude photorealism, or you risk a result that mixes the two looks.

Kling's anime prompt blog makes the same point from the other side. It advises describing the desired 2D look in positive terms, such as clean linework, flat color, a painterly background and a readable silhouette, rather than listing what to avoid. The practical reading is simple. State what you want, concretely. Save exclusions for subtitles, music and an unwanted style, and spend the remaining words on the subject, the motion and the light, which is where the formulas above put the weight anyway.

## Camera and motion: one move at a time

Kling's camera guidance for VIDEO 3.0 is blunt: use **one main camera move per shot**, because stacking several moves can give unstable results. Its photorealism blog adds that cinematography terms such as crane, dolly, orbit and tracking bring motion and drama. The camera movement control in the app offers six basic movements (horizontal, vertical, zoom, pan, tilt, roll) plus four preset Master Shots. Kling's credit policy lists master shots among premium-only functions, so those presets need a paid plan.

Image to video has its own rule. Name one exact physical motion, state the camera behavior explicitly, and add camera movement only after a static-camera version works [3]. The image already supplies the scene, so the prompt describes motion and change over time. Kling's older image-to-video guide warns that a description far from the image may trigger a camera cut. Source images that already imply motion, such as a subject mid-pose or near wind, water or light, animate more convincingly than flat arrangements. For region-level control, Motion Brush draws trajectories for up to six elements and Static Brush pins areas in place. Our [image-to-video guide](/articles/kling-ai-image-to-video) covers frames and elements in depth.


> 💡 **Lock the camera first:** For image to video, Kling advises one exact physical motion with an explicit camera instruction, adding camera movement only after a static-camera version works.


## Dialogue, voices and on-screen text

Speech is written into the prompt itself. To get dialogue in a given language, write the line in that language inside quotation marks, and the quoted words become the speech. Kling 4.0 speaks nine languages (Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi), with accents including American, British and Indian English. Kling 3.0 and 3.0 Omni cover five: Chinese, English, Japanese, Korean and Spanish. Kling 4.0 takes voice references only as audio input, so uploaded music or effects cannot steer the track. With Voice Control, which needs Native Audio switched on, a saved voice is assigned to a character in the prompt with an at sign and the voice name.

On-screen text has moved fast. Kling says 4.0 renders legible text, emoji and logos in many languages, and its guide asks you to specify font size, alignment and display timing when you request subtitles in the prompt. Earlier guidance is warier: Kling's TikTok page says text generated inside a scene comes out garbled and advises keeping lettering out of the frame. Whichever model you pick, Kling itself says output varies, and packaging, logos, colors and product details should be checked before a clip goes into an ad.

## Drafting with the DeepSeek helper, then iterating

### What the helper leaves out

Kling ships a prompt generator powered by DeepSeek-R1, opened from the whale icon at the bottom right of the video prompt box (interface as of August 2026), and any Kling account can reach it [4]. Type a rough idea or upload an image, and it returns three suggested prompts, each with its own Generate button. A Deep Thinking toggle adds framing and lighting detail. Its blind spot is motion. In Kling's own three test runs, all nine suggestions described the scene and none named a movement or camera behavior.

Kling also measured the cost of that gap. In its before-and-after comparison on VIDEO 3.0, the suggestion used as-is let the camera drift sideways by roughly a tenth of the frame width; adding one motion and a locked-off camera brought the drift to zero. The same blind spot applies when you start from an image, where the suggestions again describe the picture rather than what moves in it. Use the helper to draft the scene. Write the motion line yourself.


> ⚠ **The DeepSeek helper skips motion:** In Kling's own test runs, none of the nine suggestions from its built-in DeepSeek-R1 prompt helper named a movement or camera behavior, so add both by hand.


### Iterating without burning credits

A specific prompt improves the odds. It does not guarantee a result. Kling says so itself, since generation is not perfectly deterministic, and its advice when a prompt underperforms is to simplify to one motion. Its prompt-writing blog adds a working method: change one thing per iteration, so you can tell which edit helped. Kling describes Standard mode as the cost-effective mode for testing concepts, and recommends Professional mode, with 1080p output on a paid plan, for commercial work.

Each attempt spends credits, and the exact cost appears on the Generate button before you confirm. Kling does not publish a per-second credit rate for Kling 4.0 on its plans or guides. Failed generations return their credits, but a clip that renders and simply misses the brief does not, and fees are non-refundable once activated, as our [pricing guide](/articles/kling-ai-pricing) explains. Drafts on a free Basic account come out at 720p with the Kling watermark and may not be used commercially; see [whether Kling AI is free](/articles/is-kling-ai-free). Kling's terms also let it use inputs to train its models unless you revoke that by email. Keep confidential material out of prompts and uploads.


## FAQ

**How long can a Kling image prompt be?**

Kling's image generation guide says image prompts accept Chinese or English input of 500 characters or less. That is far shorter than the 8,000-token limit for Kling 4.0 video prompts.

**How do I keep the same character across several Kling prompts?**

Build an element from up to four multi-angle images or a 3 to 8 second video of one character, then reference it in each prompt. For a talking presenter, Kling suggests reusing the same source portrait rather than re-describing the person in words.

**Should the aspect ratio go in the prompt?**

No. Kling sets the ratio in the settings under the prompt box before generating, and it is a generation-time choice, not a crop. In Kling 4.0 Keyframes mode the ratio follows the first keyframe once images are uploaded.

**Can a prompt extend a clip in Kling?**

Not yet: video extension for Kling 4.0 is listed as coming soon in November 2026. Older models extend 4 to 5 seconds per pass with the same model and mode, and Kling warns that unrelated text in a custom extension prompt may cause a camera cut.


## The bottom line

Kling's prompting advice fits on an index card: a subject doing one thing in a place, one camera move, concrete words instead of mood words, and a stated role for every reference. Kling 4.0 adds a sentence-first structure, quoted dialogue in the target language and just two standard negatives. The built-in DeepSeek helper drafts scenes well but leaves out motion, so that line stays your job. Expect variation between runs, check text and product details before anything ships, and treat each new prompt as a test rather than a recipe.



## Sources

[1] Kling AI user guide: text to video prompt guide — https://kling.ai/quickstart/text-to-video-prompt-guide (2026-11)
[2] Kling AI user guide: kling 4 model user guide — https://kling.ai/quickstart/kling-4-model-user-guide (2026-11)
[3] Kling AI topic page: image to video prompts — https://kling.ai/topics/image-to-video-prompts (2026-11)
[4] Kling AI topic page: ai video prompt generator — https://kling.ai/topics/ai-video-prompt-generator (2026-11)
[5] Kling AI membership plans (US) — https://kling.ai/app/membership/membership-plan
[6] Kling AI Terms of Paid Service — https://kling.ai/docs/payment-policy
[7] Kling AI release notes — https://kling.ai/release-note/release-history