A useful Kling video prompt starts with what is on screen, what it does, and where it is. Add framing, camera movement, lighting, or mood when they help explain the shot. For multi-shot stories, describe the beats in order instead of cramming unrelated actions into one sentence. Treat this as a writing framework—not magic syntax or a promise that every detail will appear exactly as written.
This guide includes an example for the Kling 3.0 text-to-video API. It does not assume that prompt guidance or settings for Kling 3.0 apply unchanged to other models or providers. The API example demonstrates task submission only; callback configuration and a complete result-retrieval path are not established here.
A practical Kling prompt template
Use this as a starting point, then leave out details that do not matter to the shot:
Subject: [main person, object, or scene]. Action: [visible movement or behavior]. Setting: [place and important surroundings]. Framing/camera: [shot size, angle, or camera movement]. Lighting/mood: [visible light and atmosphere].
Kling’s published text-to-video guide treats subject, movement, and setting as the fundamentals; camera language, lighting, and atmosphere are optional additions. A later Kling-published guide uses a similar subject/action/setting/camera/lighting-and-mood framework. Neither presents it as required API syntax.
| Prompt element | What to specify | Example |
|---|---|---|
| Subject | The main focus and distinguishing visible details | A small amber glass bottle with a black cap |
| Action | What changes on screen | A bead of water slides down the bottle |
| Setting | Where the action takes place and key surroundings | On a dark stone pedestal in a quiet studio |
| Framing or camera | How the viewer sees the subject; add movement if useful | Close-up; the camera slowly pushes in |
| Lighting or mood | The visible atmosphere | A narrow warm highlight against a deep blue background |
Keep actions observable. “The bottle looks luxurious” is an impression; “a warm rim of light traces the glass” gives a visible direction. Avoid piling on conflicting instructions—for instance, asking for a locked-off shot and a fast orbit at the same time.
Build a prompt from a simple scene
Start with the scene’s essentials:
A glass perfume bottle sits on a stone pedestal.
Add details that clarify what the viewer should see:
A small amber glass perfume bottle with a black cap stands on a dark stone pedestal in a quiet studio. A bead of water slides down the glass.
Then add a deliberate shot and mood:
Close-up of a small amber glass perfume bottle with a black cap on a dark stone pedestal in a quiet studio. A bead of water slides down the glass. The camera slowly pushes in. A narrow warm highlight traces the bottle against a deep blue background.
These are illustrative rewrites, not reported generation tests. The added details make the intended scene more explicit; they do not guarantee the model will render every feature or movement exactly.
Describe motion, framing, and shot sequences
Prefer concrete directions over a vague request like “make it cinematic.” You might specify a close-up, wide shot, low angle, slow push-in, pan, tilt, or tracking shot. Kling’s published guidance gives these as plain-language prompt examples, but does not promise exact camera execution. Framing language and camera-motion controls are also not necessarily the same thing—especially when moving from a prompt to an API request.
For a short scene with distinct beats, label each shot and give it one job. For example:
Shot 1 — Wide: A cyclist enters a rain-slick street at dusk. Headlights glow in the mist.
Shot 2 — Medium tracking shot: The camera follows beside the cyclist as they pedal past closed storefronts.
Shot 3 — Close-up: The cyclist stops beneath a streetlamp; rain runs from the helmet brim.
This is an example prompt, not a claim that every Kling workflow supports this exact label format. For the official Kling 3.0 API, use the endpoint’s documented multi-shot format and constraints described below; do not assume another provider’s instructions apply.
Refine a prompt when a result misses
Use a small revision loop:
- Name the visible mismatch. For example: “The camera stayed wide; I wanted the bottle to fill more of the frame.”
- Change the part that addresses it. Rewrite “The camera slowly pushes in” as “Tight close-up; the camera slowly pushes in until the bottle fills most of the frame.”
- Keep other directions steady for the next try. That makes it easier to tell what you changed.
- Record the prompt and settings used. Save useful versions for similar scenes.
This is a practical way to organize iteration, not a proven Kling-specific optimization method. Kling’s published guide recommends trying variations in movement or framing and recording useful combinations; the supplied guides do not provide controlled comparisons or guarantee that a particular rewrite will fix a result.
Submit a prompt through the official Kling 3.0 API
The example below targets Kling 3.0 text-to-video on Kling’s own API, not fal’s separately documented API or a UI workflow. Kling’s API quick start describes creating an API key, sending it as a Bearer token, and using the api-beijing.klingai.com API domain. The Kling 3.0 text-to-video reference documents the endpoint and request fields.
Set your API key in an environment variable rather than pasting a secret into source code:
export KLING_API_KEY="your_api_key"
Then submit a five-second vertical draft:
curl --location 'https://api-beijing.klingai.com/text-to-video/kling-3.0' \
--header "Authorization: Bearer ${KLING_API_KEY}" \
--header 'Content-Type: application/json' \
--data '{
"prompt": "Close-up of a small amber glass perfume bottle with a black cap on a dark stone pedestal in a quiet studio. A bead of water slides down the glass. The camera slowly pushes in. A narrow warm highlight traces the bottle against a deep blue background.",
"settings": {
"resolution": "720p",
"aspect_ratio": "9:16",
"duration": 5
}
}'
The API reference says prompt is required, with a maximum length of 3,072 characters and a recommendation to stay at or below 2,500 characters. The example uses documented values for resolution (720p), aspect ratio (9:16), and duration (5 seconds). The endpoint reference also lists 1080p and 4k resolution, 16:9 and 1:1 aspect ratios, and durations from 3 to 15 seconds. Confirm the reference for the exact endpoint you deploy.
The create-task response includes a data.id task ID and a data.status. The documented status values are submitted, processing, succeeded, and failed. Submission alone does not return the finished video.
Callback setup and video retrieval are not confirmed here
The callback protocol documentation describes status and output fields, including a video item with a url, and says generated video URLs are cleared after 30 days. However, I could not confirm the applicable callback configuration field or its nesting in the create request for this endpoint. The submission example above therefore does not configure a callback, and this guide does not establish how to receive or retrieve the completed video. Check the endpoint reference and callback protocol before implementing result delivery; do not assume a callback field name or location.
Multi-shot through this API
The official Kling 3.0 API reference documents a multi-shot prompt format: 镜头 n, m, words; 镜头 n, m, words; with half-width punctuation between entries. Here, n is the shot number, m is its duration, and words is that shot’s prompt. The reference specifies up to six shots, a minimum of one second per shot, a total of shot durations equal to the requested video duration, and a maximum of 512 characters for each shot’s prompt. The reference also says that setting settings.multi_shot to false prevents multi-shot generation even when the prompt uses the multi-shot format; this guide does not assume a default value for that setting.
Do not automatically paste the English storyboard example above as if its labels were the documented API syntax. Format the shot entries for the endpoint you are actually calling and make their durations add up to the requested total.
What not to assume
- A prompt-writing article is not an API specification. Do not turn UI controls into JSON fields unless the API reference documents them.
- Kling models and providers can expose different settings. Guidance from another version or provider is not a substitute for the official endpoint schema.
- A prompt describes intent, not guaranteed output. The guides do not establish that wording alone will preserve every identity, exact object count, complex physical action, or camera move.
- The example above is specifically for Kling 3.0 on Kling’s official API. Verify the selected model, endpoint, field names, and limits before adapting it to another Kling version or provider.
Quick checklist
- Can you identify the subject, visible action, and setting?
- Are the movement and camera directions concrete and compatible?
- Does each detail help communicate something important in the shot?
- If there are multiple story beats, have you separated and ordered them?
- Does the API request use the correct provider, model endpoint, and documented fields?
- Have you recorded the prompt and settings so you can compare revisions?
- Have you verified callback configuration separately before implementing result delivery?
Sources
- Kling AI, Text-to-Video prompt guide — prompt structure and scene-writing advice.
- Kling AI, Kling VIDEO 3.0 prompt guide — camera language, motion, and workflow guidance.
- Kling AI, API quick start — API key and authentication setup.
- Kling AI, Kling 3.0 text-to-video API reference — request schema and multi-shot details.
- Kling AI, Callback protocol — callback payload and generated-output fields.
The API documentation is the basis for the production submission example; prompt-writing articles inform the creative guidance. Product settings and availability can change, so use the reference for the exact model endpoint you deploy.
Sources
Kling AI Text to Video Prompt Guide