Kling 3.0 Prompt Guide: Structure, Examples and Real Clips

Kling's 3.0 guide builds a prompt from six parts: subject, action, scene, camera, lighting and mood. For several shots, label each shot with its length, and for dialogue write Name (tone): line. We ran four prompts on October 28, 2026, in std mode at 5 seconds: the shot labels cut at 1.96 seconds, the dialogue matched the script in three of four transcripts, and the camera barely moved in the image-to-video run.

The Kling 3.0 prompt structure

Kling's prompt guide blog (August 7, 2026) lists the parts in this order: subject, visible action, scene, camera language, lighting and mood. Kling calls this a writing framework, and no 3.0 page makes the order mandatory.

• Shots. With automatic multi-shot on, the model reads the scene and picks the angles and compositions itself. Custom multi-shot takes a label and a length for each shot, up to six in Kling's app. With multi-shot off, the model makes a single shot by default.
• Dialogue. Write Name (tone): line and put the line in quotation marks, as the dialogue prompt below does. Kling lists five dialogue languages (Chinese, English, Japanese, Korean and Spanish) and translates any other language into English.
• Image-to-video. Describe the motion and the ending, not what the frame already shows. An older Kling page (November 24, 2025, written for 2.x) covers this too. The 3.0 guide's advice is to bind the subject as an element, so it holds steady when the camera moves.
• Start and end frames. Kling advises similar first and last frames, because large differences can trigger a shot switch. That page is dated November 24, 2025, so treat it as older advice.

Kling's 2.x formula, still on its text-to-video prompt guide, is subject, movement, scene, camera, lighting and atmosphere (November 24, 2025). We kept it apart and tested only the 3.0 form.

How we made the clips

• We ran kling-3.0 through OmniaKey on October 28, 2026, in std mode, with duration 5 for all four prompts.
• The first three were text-to-video at 16:9, output 1280×720. The fourth was image-to-video with no aspect ratio, since its first frame sets the shape; output 1176×780.
• The first frame of the fourth prompt is a GPT Image 2 still at 1536×1024, medium quality. Renders took between about 4.5 and 8.5 minutes each.
• We read the frames at fixed times, ran a scene-cut detector, measured file length and loudness, and had four audio models transcribe the speech. A line is quoted only where three of them agree.
• OmniaKey sells API access to Kling 3.0, so we ran the clips through it. The prices below are Kling's own app prices, not ours.
• Sources: Kling's VIDEO 3.0 user guide (February 6, 2026) and prompt guide blog (August 7, 2026), checked October 28, 2026.

We ran one clip per prompt, so treat each result as one sample.

Four prompts with real clips

These four prompts ran exactly as written.

1. Five parts in one paragraph: a violinist under an archway

The first prompt puts all six parts into one paragraph, with a sound line at the end.

A young street violinist in a long camel coat plays a fast, bright melody under a stone archway in an old European city at dusk. She sways with the music as a few pigeons take off from the cobblestones behind her. The camera slowly pushes in from a medium shot to a close-up of her face. Warm golden streetlights, light haze, long shadows on wet cobblestones. Sound: the violin echoing under the arch and the flutter of pigeon wings.

Settings: text_to_video, mode std, 5 seconds, 16:9, 1280×720.

Result: one continuous shot, with the camera moving in only part of the way.

• The cut detector found no cut across the 5.04 seconds.
• She plays under the archway at dusk, with a lit street behind her. At about 3.0 seconds a few pigeons lift off the cobblestones.
• By 4.9 seconds she fills more of the frame, but her face is not a close-up by the end.
• The track averages −23.2 dB. The four audio models disagreed about what plays, so we do not name any instrument.

What to change:

• Try naming the framing and the second it should arrive, for example "a close-up of her face by 4 seconds".

2. Shot labels: a lighthouse in two shots

The second prompt asks for two labelled shots, 2 seconds and 3 seconds, with the sound line last.

Shot 1 (2s): Wide shot, a lighthouse keeper in a dark wool sweater climbs a spiral iron staircase inside a lighthouse at night, holding a glowing lantern. Shot 2 (3s): Close-up, he turns a brass key and the great lamp above him flares to life, its beam sweeping across his face. Sound: his boots on the iron steps, then a deep mechanical hum as the lamp starts.

Settings: text_to_video, mode std, 5 seconds, 16:9, 1280×720.

Result: both shots appeared, with the first cut close to its label and an extra cut inside the second shot.

• Shot 1 shows the keeper in a dark wool sweater climbing the iron staircase with a lantern, clear at 1.0 seconds.
• The first cut came at 1.96 seconds, against the 2-second label. The second shot opens with his hand on a brass key by 2.4 seconds.
• A second cut came at 2.88 seconds, inside the shot labelled 3 seconds. He keeps turning the key until about 3.5 seconds, and by 4.8 seconds the lamp is lit, its beam crossing the frame as he looks up.
• The track averages −24.5 dB with no speech. The audio models described footsteps and a mechanical sound.

What to change:

• Try giving each shot one purpose, as Kling advises, with one action in each label.

3. Dialogue with a tone for each speaker

The third prompt is the dialogue test: two speakers, each with a tone in parentheses and a line in quotation marks.

Inside a small bookshop on a rainy afternoon, an old shop owner with round glasses hands a wrapped book across the wooden counter to a teenage girl in a yellow raincoat. Medium two-shot, then the camera slowly pushes in on the girl's smile. Owner (warmly, in a slow, deep voice): "Bring it back when you've finished it." Girl (excited, speaking quickly): "I'll read it tonight!" Rain patters on the shop window.

Settings: text_to_video, mode std, 5 seconds, 16:9, 1280×720.

Result: no cut, and the dialogue matched the script in three of four transcripts.

• The owner hands the wrapped book across the counter to the girl at 0.3 and 1.7 seconds.
• The camera moves on to the girl at 3.3 seconds, and by 4.9 seconds her face fills the right half of the frame.
• Three of the four audio models transcribed the same two lines: "Bring it back when you've finished it. I'll read it tonight!" The fourth dropped the contraction in "you've".
• The track averages −26.0 dB.

What to change:

• Keep this shape for the next two-person scene: a tone in parentheses after each name, then that speaker's line in quotation marks.

4. Image-to-video: describe the motion

The fourth prompt starts from a still we made with GPT Image 2 at 1536×1024: a woman in a red scarf reading on a park bench in autumn, photographed from her left side. The still's prompt asked for a realistic 35mm photo at eye level, in golden light with fallen leaves on the path.

The camera slowly arcs around to the front of the woman, who closes her book, lifts her head and smiles warmly at the camera. A few leaves drift down around her.

Settings: image_to_video, mode std, 5 seconds, first frame 1536×1024 with no aspect ratio, output 1176×780.

Result: she turns and smiles, but the camera barely arcs and the book stays open.

• At 1.5 seconds she is still reading. At 3.0 seconds she has turned toward the camera and is smiling, with the book still open in her hands.
• At 4.9 seconds the book is still open and the camera has barely moved. The turn and the smile happened; the arc and the closed book did not.
• The prompt has no sound line, and the track is near-silent: an average of −50.0 dB, peaking at −20.9 dB, with no speech.

What to change:

• Try stating the end state in one line, for example "the book closed by the end", and keep the camera to one move.
• Add a sound line, as the first two prompts did.

Common failures and fixes

These come from the four prompts above. The section numbers match the headings.

What a clip costs

Kling's own app charges credits per second. The rate depends on the resolution and on whether native audio is on.

These figures come from Kling's own app, checked October 28, 2026. The 5-second column is the per-second rate times five, and it matches Kling's example of 60 credits for 5 seconds of 1080p with native audio. Voice control adds 2 credits per second, which makes that same clip 70 credits. The 4K mode costs 30 credits per second, according to Kling's prompt guide blog (August 7, 2026).

Credits are not dollars, so compare models on each vendor's own pages. For a side-by-side, see our Kling 3.0 vs Wan 3.0 comparison. For OmniaKey's current rates, see the Kling 3.0 model page.

Run these prompts through the API

This request sends the lighthouse prompt from section 2 as one string. Its idempotency key keeps a retry from creating a second job.

curl https://api.omniakey.com/v1/media/tasks \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: lighthouse-shots-001" \
  --data-binary @- <<'JSON'
{
  "model": "kling-3.0",
  "type": "video",
  "operation": "text_to_video",
  "input": {"prompt": "Shot 1 (2s): Wide shot, a lighthouse keeper in a dark wool sweater climbs a spiral iron staircase inside a lighthouse at night, holding a glowing lantern. Shot 2 (3s): Close-up, he turns a brass key and the great lamp above him flares to life, its beam sweeping across his face. Sound: his boots on the iron steps, then a deep mechanical hum as the lamp starts."},
  "parameters": {"mode": "std", "duration": 5, "aspect_ratio": "16:9"}
}
JSON

We sent no multi_shot field, as in our test: both shot labels sit inside input.prompt.

1. The reply is 202 with a task id.
2. Fetch https://api.omniakey.com/v1/media/tasks/{id} with that id and loop on status until it is final. Renders took about 4.5 to 8.5 minutes, so keep polling rather than resubmitting. Replaying the same key and body returns the same task.
3. Download the clip from assets[0].url. That link is signed and lasts about an hour, so poll again for a fresh one if it expires.

For image-to-video, set operation to image_to_video and send the first frame in input.media with type first_frame, as an HTTPS URL or a data URI. last_frame is optional, but the first frame is required. aspect_ratio is rejected there, and the whole request body is capped at 1 MiB.

OmniaKey takes duration 5, 10 or 15 only, while Kling's own API takes whole seconds from 3 to 15. Start from the Kling 3.0 model page, and see the video API docs.

Common questions

How long can a Kling 3.0 prompt be?

Kling's API caps prompt at 2,500 characters for text-to-video, and the same cap applies to image-to-video. Turbo allows 3,072 and recommends staying under 2,500, so quote the cap for each endpoint rather than one number for the model.

Does Kling 3.0 support negative prompts?

The API reference recommends putting negative sentences inside the positive prompt, as well as using its negative prompt field. None of the 3.0 guides gives negative-prompt advice, so treat that line as API guidance.

Can I get a 3-second or 8-second clip?

Kling's own API takes whole seconds from 3 to 15, so 3 and 8 both work there. OmniaKey takes only 5, 10 or 15 seconds.

What is the Turbo format?

Turbo takes one inline string, written shot n, m, words;, with each shot separated by a semicolon. Turbo is a separate endpoint, and this format is not how 3.0 prompts are written.

Update log

• October 28, 2026: first version, four prompts run the same day.