A Kling prompt guide built from Kling's own Kling 4.0 manual. It covers prompt structure, length limits, @ references, dialogue, keyframes, prices and examples.
Method. Kling 4.0 user manual, prompt guide, API index and pricing page read November 18, 2026; Kling's two sample prompts are quoted, and our three examples are untested illustrations. Vinc makes an AI film studio.
Kling prompt guide: the short answer
The Kling 4.0 Video User Manual sets the order: "Start by writing the subject, scene, action, and emotion in a single sentence. Then add the camera work, timing, audio, and constraints that genuinely affect the result."
Name each uploaded image or video with an @ tag and say what role it plays. Write dialogue in quotation marks in the language it should be spoken in, and write "no subtitles" or "no background music" to switch those off. For how Kling's rules compare with Google's, Runway's and LTX's, see How to Prompt AI Video Models.
Which Kling version does this guide cover?
Kling 4.0 and Kling 4.0 Flash. Kling's API documentation index lists 4.0 as its newest video model on 18 November 2026, with 3.0, 3.0 Omni, 2.6 and 2.5 Turbo still available. From the user manual:
Kling 4.0
Kling 4.0 Flash
Clip length per generation
3–30 s, in 1-second steps (default 10 s)
3–20 s
Resolution
720p, 1080p (default); 4K and 4K HDR "coming soon"
Two-channel stereo, audio-visual sync on by default
Same
Kling says extension up to 2 minutes is "coming soon". There are three API endpoints:
Text-to-Video, for when "you only have text".
Image-to-Video, for when "key visual moments are already defined".
Omni Reference, which combines images, videos, elements and voice profiles and "classifies the intent of the user's prompt and routes it to the appropriate pipeline".
Smart shot-splitting is on by default: the model organises shots and pacing itself, which suits a 30-second multi-shot clip but surprises anyone expecting one continuous take.
What is the Kling prompt structure?
Kling publishes the structure as a fill-in template in section 8.2 of the manual:
[Asset Roles] @Image1 is ___, keep its ___ consistent. @Image2 is ___, keep its ___ consistent. Reference the ___ of @Video1.
[Overall Settings] Scene and texture: ___. Visual style: ___. Camera tone: ___.
[Sequence of Events] 0–X s: ___ (visuals + action + emotion) X–Y s: ___ (camera movement + additional details) Y s–end: ___ (closing shot)
[Audio] Who says what or narration: ___. Ambient sounds: ___. Music: ___. Sound effects: ___.
The four blocks of Kling's template (user manual, section 8.2) in its order. Skip any block you don't need; a text-only prompt has no asset roles.
The manual also shows the opposite extreme, "One-Sentence Generation": given only "An old man ranting about AI.", the model "will expand the simple prompt by filling in the storyline, scene blocking, and camera movements". When your description is clear enough, it "follows it faithfully without adding its own interpretation". A short prompt hands the creative decisions to Kling; a detailed one keeps them with you.
Describe what is visible, not adjectives. The manual's first rule is to "Avoid abstract terms":
Not recommended
Recommended (Kling's wording)
Premium feel
Side backlighting outlines the object's edges, the background is blurred, and fine matte grain is visible on the surface
Vintage feel
Yellowed walls, worn furniture edges, and warm incandescent lighting
A middle-class family's kitchen
Wood-grain cabinets, a rice cooker and seasoning jars on the countertop, sticky notes on the refrigerator door
Don't show the element's face while they're on the phone
Show the element in profile while talking on the phone
The limit is 8,000 tokens, which the API expresses as up to 26,000 English characters or 11,000 non-English characters. Mixed input must satisfy (E = English characters, N = non-English):
E / 26000 + N / 11000 <= 1
Asked "Are longer prompts better?", Kling's FAQ answers "Not necessarily." Its examples run from one sentence to several hundred words, with long ones reserved for 15- and 30-second multi-shot pieces.
How do you reference images and videos with @ tags?
Uploaded assets are numbered in upload order (@Image1, @Image2, @Video1), and "Every asset must have its role declared"; otherwise "the model will infer each asset's intended use on its own". Kling's example:
@Image1 defines the scene. Preserve its spatial layout, furnishings, and lighting. @Image2 defines the female protagonist. Preserve her face, hairstyle, and clothing. The female protagonist walks toward the window in the scene, stops, and looks back.
When assets supply different things (character, action, camera, audio), give each one's scope and "explicitly state what should not carry over". Omni Reference takes 15 assets per task: up to 10 images (30 MB each), 5 videos totalling 30 seconds, 7 elements and 3 voice profiles (5–30 seconds of source audio each). A reference video can supply only the camera path, a way to borrow a move from your own footage or a previs pass: "Reference the camera path and pacing of @video_1. Use the car in @image_1 as the subject."
How do you write dialogue and audio in Kling?
Kling says to "write the dialogue in that language directly inside quotation marks", optionally with age, vocal tone, speech rate and emotion. The manual lists Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi, plus Sichuanese and Cantonese dialects and American, British and Indian English accents. Its multilingual example ties each line to an action:
1. Tourist raises her map toward the interpreter:
"¿Dónde está la estación de metro más cercana?"
Audio rules from the manual:
No music or sound effects upload. "Audio input is supported only as voice profiles." Music and effects "should be described in the prompt — the model will generate them from the visuals".
Sync is always on, "with no toggle to turn them off".
Switch off defaults. "The model may add subtitles and background music by default." Write "no subtitles" or "no background music".
Google's Veo, by contrast, avoids quotation marks so lines aren't rendered as on-screen text (see our Veo 3 Prompt Guide). For on-screen text you do want, Kling says to "Specify the exact text content, position, alignment, on-screen duration, and whether it should sync with spoken dialogue."
Camera language in Kling prompts
The manual names push-in, pull-out (crews say dolly in and out), pan, lateral camera movement, tracking shots and orbit shots. The prompt guide's examples get more precise: "One uninterrupted lateral tracking shot", "subtle 3% push-in", "Pull focus onto her face with a small correction", "Hard cuts between poses; no morphing". To stop the model cutting, ask for one continuous master shot "with no visible cuts". On-set meanings are in Film Terms: Camera Movement and Camera Language for AI Video Prompts.
Do Kling prompts need negative prompts?
Kling 4.0 has no separate negative field in the endpoints we read; the text-to-video API reference says the one prompt field "can include positive and negative descriptions". The manual says "Prefer positive descriptions of what you want to see", allowing targeted exclusions (for a stylised piece, "explicitly exclude realistic texture") but "Don't expand the negative list into a long register of items unrelated to the task."
First frame, last frame and keyframes
The manual's rule: "For the first frame, write out what happens next; for the last frame, write out the process that leads up to it." For intermediate keyframes, describe "what happens at that node and how the camera arrives at that position" in time segments:
The prompt guide's patterns are "@Image1 as the first frame, [what happens next]" and "@Image1 as the first frame, @Image2 as the last frame, generate [the process in between]", the second suited to transformations and before-and-after shots. If you reference a storyboard sheet, ask for finished frames, or "grid lines, title bars, or Chinese annotations from the storyboard sheet may appear" (see Storyboards, Animatics, Shot Lists and Overheads).
Editing an existing video: "Edit", not "Reference"
Name the video, the change, and what must stay the same:
Edit @Video1. Change ___ to ___. Keep everything else unchanged.
The verb matters. "Using only 'Reference' may route the request into reference-based generation instead of editing."
Kling prompt examples
The first is from Kling's documentation; the second is our untested illustration of the rules above.
From Kling's docs: aspect ratio and duration in the prompt (with aspect_ratio set to adaptive)
Generate a 15-second, 9:16 vertical video: morning at the seaside, a seagull skims across the water, camera slowly pushes forward.
Ours: a single dialogue shot
A tired night-shift nurse leans against a vending machine in a hospital corridor at 3 a.m., relieved and half-smiling.
Camera: medium close-up, eye level, slow push-in.
0–4 s: she presses a button and waits. 4–8 s: a can drops; she laughs quietly.
Audio: fluorescent hum, distant monitor beeps. In a soft, hoarse voice, slowly, she says: "Finally, something works."
No subtitles. No background music.
What does Kling cost per second?
Kling publishes API prices with dollar equivalents on its video pricing page. Without an input video, Kling 4.0 bills output seconds only; with one, it bills input plus output, or output × 2, at a lower unit rate.
Kling list price per second at 1080p (no input video)
Kling 4.0$0.235
Kling 3.0 (audio)$0.168
Kling 3.0 Turbo (audio)$0.140
Kling 4.0 Flash, 720p$0.094
Kling API video pricing, checked 18 Nov 2026. Kling 4.0 Flash is 720p only. Kling 3.0 is shown with native audio; without audio it is $0.112.
By our arithmetic, one attempt at a 10-second 1080p Kling 4.0 clip costs $2.35, and 30 seconds costs $7.05. Runway resells Kling 4.0 at the same $0.235 list price, discounted to $0.188 until 30 November 2026 (API pricing). Vinc (us) offers Kling 3.0 (v3 Pro and v3 4K): about 13.4 credits a second without audio and 20.2 with, roughly $0.13 and $0.20 at Starter's $0.01 a credit (Vinc pricing). Other models are in AI Video Models Compared.
How we compiled this
On 18 November 2026 we read Kling's own manual, prompt guide, API docs and pricing, plus Runway's API pricing; every quoted rule comes from those pages. The first example is Kling's; the second is our untested arrangement of its rules. We left out third-party "Kling prompt" lists, many of which still describe Kling 1.x–2.x behaviour. Vinc makes an AI film studio.
AI Music for Film: Suno, Udio and Other Generators on Licences, Ownership, Lawsuits and Cue Sheets
Whether an AI-generated track can go in your film depends on the plan you made it on, whether the terms exclude film and TV, and whether anyone can own it at all. This guide compares the published terms of Suno, Udio, ElevenLabs Music, Stable Audio and AIVA, sets out the label lawsuits as reported, and explains what US copyright law and North American PRO rules mean for your cue sheet.