Free Kling 3.0 Prompt Generator

TL;DR: Paste the generator block below into any chat model, give it your idea, a duration between 3 and 15 seconds, and your dialogue. It returns a Kling 3.0 shot list with per-shot durations that sum to your budget, a timecode version, a negative prompt, and the settings to run it. Everything below it is sourced and dated.

What is a Kling prompt generator?

A Kling prompt generator is anything that turns a one-line idea into a prompt Kling will actually follow. Most of the hosted ones are a text box in front of a language model with a prompt you never get to see. That is fine until the output is wrong and you have no idea which instruction to change.

So here is the prompt itself. Copy it, keep it, edit it. It is built around the one constraint that governs Kling 3.0 work: a single generation runs 3 to 15 seconds, and if you are using multiple shots, their durations have to add up.

You are a Kling 3.0 shot planner. Turn my idea into a Kling VIDEO 3.0 prompt.

MY INPUT
Idea: [one or two sentences]
Total duration: [a whole number of seconds, 3 to 15]
Aspect ratio: [16:9 | 9:16 | 1:1]
Dialogue: [none, or the exact lines with who says each one]
Look: [two or three reference words, e.g. handheld documentary, overcast, 35mm]

RULES
1. Total duration is a hard budget. Emit between 1 and 6 shots. Every shot
   duration is a whole number of seconds and they must sum to exactly the total.
2. A shot containing dialogue gets at least 2 seconds. Budget roughly one second
   per six spoken words, plus a beat at each end.
3. Write each shot on one line, in this shape:
   Shot N (Xs): [framing and angle], [subject and what changes], [camera move],
   [light and colour].
4. Put a spoken line inside the shot that shows the speaker, as:
   Name (delivery): "line". Delivery is a tone word, plus an accent if it matters.
5. Every shot must name something that changes during it. No frozen images.
6. Keep the whole prompt under 2,500 characters. Cut adjectives before you cut shots.
7. Do not invent brand names, on-screen text or logos unless I gave them to you.

OUTPUT, in this order, and nothing else
A. SHOT LIST: one shot per line in the Shot N (Xs): format.
B. TIMECODE VERSION: the same shots as [00:00 - 00:0X] ranges.
C. NEGATIVE PROMPT: one comma-separated line, limited to artefacts this
   specific idea is likely to produce.
D. SETTINGS: duration, aspect_ratio, generate_audio (true/false),
   shot_type (customize/intelligent).
E. BUDGET CHECK: the shot durations added up, confirming they equal the total.

Section E is not decoration. Shot durations that do not sum to the requested total is the single most common failure in generated Kling prompts, and a model that has to show its arithmetic gets it right far more often than one that does not.

If you would rather not put a model in the loop at all, this is the same structure as a fill-in template:

Shot 1 (__s): [wide/medium/close], [angle], [subject + what changes],
              [camera move], [light + colour].
Shot 2 (__s): ...
Shot 3 (__s): ...

Negative prompt: [artefacts specific to this shot type]
Settings: duration __s · aspect __ · audio on/off · shot type customize

What Kling 3.0 actually accepts

Direct answer: Kling 3.0 takes a prompt of up to 2,500 characters, a duration of 3 to 15 whole seconds, an aspect ratio of 16:9, 9:16 or 1:1, an optional per-shot multi_prompt array, a negative prompt, a CFG scale, and an audio toggle. There is no frame rate parameter and no published frame rate.

Two of those deserve a second look. duration is a string, so 5 fails where "5" works, which catches out roughly everyone on their first API call. And generate_audio defaults to on, so silent b-roll costs you audio money unless you explicitly turn it off.

How do you write a timed shot list for Kling 3.0?

Kling documents two shapes for multi-shot prompts, and both appear in its own examples. Use either. Do not mix them in one prompt.

Labelled shots with durations. This is the shape Kling's VIDEO 3.0 Omni guide uses throughout:

Shot 1 (3s): Mid-shot. A neon-lit street corner late at night, wet pavement
reflecting light. A man in a long coat leans against a phone booth, smoking,
strong motion blur behind him.
Shot 2 (2s): Cut to close-up. His profile half-hidden in shadow. He looks down
and says, "You still haven't decided which road to take?"
Shot 3 (4s): Cut to close-up of a woman, lips and swaying earrings. She flips a
coin and says, "I heard there's a place where people never ask for directions."

Timecode ranges. Kling's guide also carries a partner example built entirely from timecodes, with an explicit audio line per beat:

[00:00 - 00:02] Medium shot: a man gestures emphatically with a lit cigarette,
walking towards a row of lockers, smoke curling around his hand.
Audio: the faint crackle of the cigarette tip under his words.
[00:02 - 00:04] Close-up: his weathered face fills the frame, eyes wide, jaw
working as he speaks.
Audio: crackle continues, room tone low and tight.
[00:04 - 00:06] Cutaway: a young woman looks down at her taped hands, stoic,
refusing to react.
Audio: crackle softens, her breath barely audible.

Timed beats inside one continuous take. If you want a single unbroken shot rather than a cut sequence, Kling's own 15-second example directs the clock instead of the cut: "At the 4th second, the camera accelerates forward with her... At the 8th second, the camera gradually zooms in to a medium shot... In the final 3 seconds, the camera keeps moving forward." That is a genuinely different tool from Multi-Shot and it is worth knowing about, because a long take is often the more convincing choice for a 15-second clip.

For dialogue, Kling pairs the speaker with the line and puts the delivery in brackets: Mom (softly, in a surprised tone): Wow, I didn't expect this plot at all. Its guide documents Chinese, English, Japanese, Korean and Spanish, plus Chinese dialects such as Cantonese and Sichuanese and English accents including American, British and Indian. Anything outside those five languages gets translated to English. If you are coming from Google's model, the attribution habit transfers but the syntax does not, so read our Veo 3.1 dialogue guide rather than reusing Veo phrasing here.

Four worked examples

Each of these is a finished generator output. The arithmetic is done, and the shot count is inside the documented six-shot ceiling.

A 15-second product spot, four shots, 16:9.

Shot 1 (4s): Wide, slight low angle. A matte black espresso machine on a
concrete counter in an empty café at dawn. Steam begins to rise. Slow dolly in.
Cold blue window light with one warm practical behind.
Shot 2 (3s): Macro. A single stream of espresso falls into a glass cup, crema
building. Camera locked, rack focus from spout to crema.
Shot 3 (4s): Medium, over the shoulder. A barista's hands lift the cup and turn
it a quarter turn. Handheld with a small drift. Warm key from the left.
Shot 4 (4s): Wide again, mirrored from shot 1. The barista steps out of frame,
the café lights come up, the machine sits alone. Slow dolly out.

Negative prompt: warped hands, extra fingers, liquid flowing upward, floating
cup, morphing logo, text artefacts, jittery camera
Settings: duration 15s · aspect 16:9 · audio on · shot type customize
Budget check: 4 + 3 + 4 + 4 = 15

A 12-second two-hander with dialogue, three shots, 16:9.

Shot 1 (4s): Medium two-shot. Two colleagues at a small kitchen table in a
flat, late evening. Overhead pendant light, warm. He slides a folder across.
Camera static.
Shot 2 (4s): Close-up on her. She reads for a beat, then looks up.
Sara (flat, tired, British accent): "You already signed it."
Shot 3 (4s): Close-up on him, slightly wider than hers. He does not look away.
Tom (quiet, unapologetic): "I signed it on Tuesday."

Negative prompt: mismatched eyelines, mouth moving with no sound, duplicated
person, warped face, subtitles, watermark
Settings: duration 12s · aspect 16:9 · audio on · shot type customize
Budget check: 4 + 4 + 4 = 12

Notice how short those lines are. Four words and five words, in four-second shots. Overrunning the clock is what makes Kling rush a delivery, and the generator's one-second-per-six-words rule exists to stop you writing a paragraph and hoping.

A 9-second vertical social cut, three shots, 9:16.

Shot 1 (3s): Close-up, phone-height. A pair of running shoes hitting wet
tarmac, water spraying. Handheld, following. Overcast grey, desaturated.
Shot 2 (3s): Medium, low angle tracking backwards. A runner in a black shell
jacket, breathing hard, headphones in. Rain on the lens.
Shot 3 (3s): Wide, static. The runner crests a hill and stops. City below,
lights coming on. Camera holds.

Negative prompt: sliding feet, feet not touching ground, extra limbs, warped
face, rain frozen in place, text
Settings: duration 9s · aspect 9:16 · audio off · shot type customize
Budget check: 3 + 3 + 3 = 9

One continuous 15-second take, no cuts, 16:9. Here the generator emits a single shot and directs the clock:

A 15-second continuous take, one unbroken shot with no cuts. A woman in a dark
green coat walks the length of a covered market at closing time, the camera
tracking beside her in a steady medium shot. Cold overhead strip lighting mixed
with warm stall lamps. At the 4th second she stops at a flower stall and the
camera stops with her. At the 8th second she picks up a single white flower and
turns it over. At the 12th second she walks on, the camera resuming beside her.
In the final 2 seconds the stall lights behind her switch off one row at a time.

Negative prompt: hidden cut, jump cut, teleporting subject, warped hands,
flickering lights, morphing background
Settings: duration 15s · aspect 16:9 · audio on · shot type customize
Budget check: single shot, 15

If you want the underlying grammar rather than finished examples, our six-part Kling prompt framework covers subject, action, context, style, camera and motion, and the filmmaker's guide to lens, light and framing covers the vocabulary the generator is filling in.

Multi-Shot or Custom Multi-Shot?

In the Kling web app these are two switches, and the relationship matters. Kling's VIDEO 3.0 guide states that Multi-Shot must be enabled before Custom Multi-Shot becomes available, and that with Multi-Shot off, the model defaults to a single-shot video. With Multi-Shot on and Custom Multi-Shot off, the model plans the transitions itself and will collapse to a single shot if it judges the scene better served that way.

Through the API the same choice is the shot_type field. customize means your shot list is the plan. intelligent means the model decides where the cuts fall.

Use customize when the cut is the point: a shot-reverse-shot exchange, a product reveal, anything where a beat has to land at a specific second. Use intelligent when you have written a scene rather than a shot list and you would rather not fight the model about coverage. The generator above defaults to customize because a timed shot list is the whole reason to use it.

Through the API, the same three shots become an array:

{
  "multi_prompt": [
    { "prompt": "Wide establishing shot of a Tokyo street at night, rain falling.", "duration": "4" },
    { "prompt": "Close-up of a neon sign reflected in a puddle.", "duration": "3" },
    { "prompt": "A lone figure walking away under an umbrella, warm streetlight.", "duration": "5" }
  ],
  "shot_type": "customize",
  "aspect_ratio": "16:9",
  "generate_audio": true
}

Note that prompt and multi_prompt are mutually exclusive. Sending both is an error, not a merge. The per-shot duration values are strings here too, and they carry their own enum running from "1" to "15", so a one-second cutaway is legal inside a storyboard even though a one-second whole clip is not.

What belongs in the Kling negative prompt?

The default is blur, distort, and low quality, which is close to useless because it describes an outcome rather than an artefact. Replace it with the specific failures your shot type invites.

Human motion invites sliding feet, feet not contacting the ground, and duplicated limbs. Hands invite extra fingers and fused fingers. Faces in close-up invite warped features and shifting identity between frames. Liquid and smoke invite frozen particles and unnatural flow direction. Any shot with a logo or sign invites text drift, because Kling 3.0 specifically markets accurate lettering and will happily generate lettering you did not ask for.

Keep it short and concrete. A negative prompt listing thirty generic quality words does less than one listing six things that actually go wrong in this shot. The field accepts up to 2,500 characters; you should be nowhere near that.

Which Kling 3.0 mode should you target?

Resolution and cost are decided by which route you call, not by a parameter inside the prompt. Kling's own developer pricing page and the fal.ai partner schema give the following picture. All figures checked August 26, 2026.

The Turbo row is the one to read carefully. Turbo is cheaper per second at 1080p than Kling 3.0 with audio, but its published schema accepts only prompt, multi_prompt, aspect_ratio and duration. No negative prompt, no CFG scale. If your workflow depends on suppressing a specific artefact, Turbo will not let you.

In the consumer web app the currency is credits rather than dollars. Kling's VIDEO 3.0 guide lists 12 credits per second for 1080p with native audio, 8 without, and 9 and 6 respectively at 720p, with voice tone control adding 2 credits per second. A 5-second 1080p clip with audio is 60 credits by that table.

Does Kling 3.0 really do 4K at 60fps?

Half of that is documented and half of it is not, and the difference matters if you are planning a delivery spec.

4K is real, and it is separate from the February launch. Kuaishou's launch release on February 5, 2026 announced Video 3.0, Video 3.0 Omni, Image 3.0 and Image 3.0 Omni, and the 2K and 4K claim in that release is attached to the image models. Kling's VIDEO 3.0 user guide, published the following day, lists 1080p and 720p only. Native 4K video arrived later: Kling's own blog states that on April 23, 2026 it launched native 4K generation for the Kling 3.0 video series, one-click and not upscaled. Kling's developer pricing now carries a 4K column at $0.42 per second for Kling 3.0 and 3.0 Omni, and a dash for Turbo. The fal.ai 4K route exists as a separate endpoint and notes that only the Singapore server supports it.

The frame rate is a different story. Kuaishou's launch release does not mention frame rate. Kling's VIDEO 3.0 and VIDEO 3.0 Omni user guides do not mention frame rate. Kling's native 4K blog post does not mention frame rate. Kling's developer pricing table does not list one, and fal.ai's schema and specification table publish no output frame rate either. The widely repeated "native 4K at 60fps" figure does not appear in any Kling or Kuaishou material this post could read.

Why your generated prompt still failed

The shots do not add up. Kling honours the total duration, so a shot list summing to 17 seconds inside a 15-second budget gets compressed somewhere, usually by rushing your last beat. Re-run the budget check.

The dialogue is clipped. The line was too long for its shot. Read it aloud with a stopwatch. If it does not finish comfortably with a beat spare at each end, cut words rather than asking for faster delivery.

You got one shot when you asked for four. In the web app, Multi-Shot was off, or it was on with Custom Multi-Shot off and the model decided a single shot suited the scene. Through the API, shot_type was intelligent. Set customize and label the shots.

The wrong character speaks. Attach every line to a visual description or a bound @Element, not to a name the model has never seen. Kling's guide is explicit that pairing the character with their line is what resolves ambiguity in multi-character scenes.

Text or a logo appeared that you never asked for. Kling 3.0 actively generates lettering. Add text artefacts to the negative prompt and remove any word from your prompt that reads like signage.

Feet slide, hands warp, the face changes. These are what the negative prompt is for. Replace the default with the six things this specific shot invites.

Prompt Architects does not render video. It builds the prompt, scores it, and keeps it: the generator above lives well as a saved template with [idea], [duration] and [dialogue] as variables, so the next shot list is a fill-in rather than a rewrite. Enhancement, video prompt tools and the prompt library are all on the free plan, and current paid pricing is on the pricing page. If you want the same treatment for Google's model, we published a free Veo 3 prompt generator as well.

Where to check all of this yourself

Model specs go stale fast, and several numbers in circulation about Kling 3.0 are already wrong. These are the pages verified on August 26, 2026, so you can re-check rather than trust this post in six months:

• Kling VIDEO 3.0 Model User Guide, February 6, 2026, for durations, Multi-Shot behaviour, dialogue syntax, languages and credit pricing: kling.ai/quickstart/klingai-video-3-model-user-guide
• Kling VIDEO 3.0 Omni Model User Guide, February 6, 2026, for storyboard control, @Element references and Omni pricing: kling.ai/quickstart/klingai-video-3-omni-model-user-guide
• Kling AI blog on native 4K, published May 20, 2026, for the April 23 4K launch date: kling.ai/blog/kling-ai-introduces-native-4k-video-model
• Kling developer pricing, for per-second API list prices by resolution: kling.ai/dev/pricing
• Kuaishou launch release, February 5, 2026, for the Kling 3.0 model line-up and the 15-second duration: globenewswire.com/news-release/2026/02/05/3232837/0/en/kling-ai-launches-3-0-model-ushering-in-an-era-where-everyone-can-be-a-director.html
• fal.ai Kling Video v3 API reference, for the parameter schema, defaults and the 1-to-6-shot storyboard limit: fal.ai/models/fal-ai/kling-video/v3/standard/text-to-video/api
• fal.ai Kling Video V3 4K page, for the 4K route, its Singapore-only note and $0.42 per second: fal.ai/models/fal-ai/kling-video/v3/4k/text-to-video

One access note, since it will affect anyone trying to repeat this. Kling's developer API reference at app.klingai.com renders entirely client-side and returns no readable content to a plain fetch, and kling.ai redirects and blocks some clients. The quickstart guides and the blog are server-rendered and readable. The parameter names, types and defaults in this post therefore come from a published API-partner schema rather than from Kling's own API reference, which is a real limitation and not one worth hiding.