Kling 4.0 Prompt Guide: Multi-Shot Storytelling

Most prompt guides are written by people who generated four videos and called it research. This one is different for a boring reason: prompting is literally our job. Every clip in the kling-4.ai gallery came out of a prompt we wrote, re-rolled, swore at, and rewrote — and in this post we're breaking down the structure that survived all that iteration.

Fair warning: multi-shot prompting is more like writing a one-page shooting script than typing a wish. If you want "cat riding skateboard, 4K", you don't need this guide (and honestly, that prompt will work fine). This is for when you want a sequence — cuts, coverage, pacing, a character who stays the same person across shots.

Why multi-shot prompting is a different sport

Old-gen models took a vibe and gave you a vibe back. Kling 4.0 takes direction. The model can hold up to 120 seconds of structured sequence with real cuts, which means the prompt has to answer questions a director would answer: what does each shot show, how long does it hold, where's the camera, what's the light doing, what do we hear.

The good news — you don't have to invent this structure. Film production solved it a century ago. A working Kling 4.0 prompt is basically a compressed call sheet.

The anatomy: six blocks

Every serious prompt we ship has the same six blocks, in roughly this order:

1. Style & image parameters. Shot on what? Phone footage or anamorphic glass? Frame rate, grain, any "do nots" (e.g. no beauty smoothing). This block sets the physics of the whole clip.
2. Scene setting. The location, dressed in specifics. Not "a bedroom" — "warm-toned bedroom, pink velvet headboard, leopard-print linens, soft yellow light from the right."
3. Character lock. Who's in it, described precisely — or better, locked to a reference image ("the woman from Reference Image 1; do not alter facial features or clothing").
4. The shot list, timestamped. The heart of it. 00:00–00:03 Shot 1: ... What happens, what the frame is, where the camera moves.
5. Grade & texture. Color treatment, film stock references, halation, flare. "Kodak Vision3 500T, 35mm grain, lime-teal tones" gets you a look one adjective never will.
6. Audio layer. Kling 4.0 generates sound natively, so tell it what to hear: room tone, foley, breath, a synth undercurrent. Prompts that skip this block get generic ambience.

Let's see it working, with receipts.

Example 1 — UGC realism (the hardest easy thing)

Everyone thinks cinematic is the hard mode. Wrong — casual is. This clip has to look like a creator filming herself on a front camera at arm's length, while pulling off four wardrobe-and-location swaps inside one unbroken take. Making that read as "real footage" means prompting against the model's instinct to make things pretty. Condensed version of the prompt behind it:

Single unbroken long take, 0:00–0:15, genuine UGC selfie video.
9:16 vertical, front camera held at arm's length. Natural hand tremor,
authentic skin texture, visible pores, natural light only. No cinematic
studio look, no beauty filters, no plastic CGI skin.

Character: the woman from Reference Image 1. Same face, same identity
across all four looks — only outfit, hair, and location change.

Effect: each finger snap swaps outfit + background in-shot, with a quick
halo and motion blur sweep. No hard cuts, no glitch effects.

Look 1, 0–4s: casual black top, cozy apartment, soft daylight.
Dialogue: "Hey guys... wanna see my little secret?" Snap at 4s.
Look 2, 4–7.5s: champagne sequined dress, rooftop at night, city
skyline bokeh. Playful shrug. Snap at 7.5s.
Look 3, 7.5–11s: Y2K street style, graffiti alley, daytime, high
ponytail, oversized leather jacket. Snap at 11s.
Look 4, 11–15s: red evening gown, opera house steps, golden hour.
Closing line: "So... which look is your favorite? Tell me."

Audio: dialogue recorded close to the phone mic; a crisp snap sound
timed to every transition.

Notice what's doing the heavy lifting: the negatives ("no beauty filters", "no cinematic studio look") and imperfection as instruction (hand tremor, visible pores, natural light only). If you don't ask for flaws, Kling gives you a shampoo ad.

The four looks the prompt specifies, pulled as frames. This is what a "character lock" instruction buys you: the identity survives every snap.

Example 2 — Product commercial pacing

Ads live and die on rhythm: hook, product beauty, human reaction, end card. The structure to steal here is shot duration as pacing — commercials cut fast early and hold long on the product. In your shot list, give the hook 2 seconds, the product macro 4. The model respects timestamped durations remarkably well, and that alone will make your output feel "edited" rather than generated.

The reaction beat. Every ad needs one, and it's a one-line instruction: "close-ups of happy people smiling as they take their first bite, natural approachable reactions."

Also: put the product in a reference image, always. Label typography is exactly the kind of detail that drifts between shots when it's only described in text.

Example 3 — Full cinematic sequence

The deep end: a scientist versus a four-meter mantis in an overgrown lab, played as a 1980s creature thriller. The full prompt is long (they get long — ours run 300–500 words for sequences like this), but the blocks that make it work:

Camera, per segment:
Seg 1: slow steady push-in on the researcher and her flashlight, then
pan up to reveal the creature. 
Seg 2: stabilized low angle for the dive under the table; rotate the
lens on a 90-degree arc as she throws the beaker.
Seg 3: seamless follow shot toward the exit; sparking wire acts as the
scene's key light; end on a slow push into a medium on her face.
Overall: 35mm anamorphic, natural motion blur, light handheld weight,
no unmotivated shake.

Grade: flickering cool fluorescents + sick green jungle ambience, deep
shadows, lime-teal treatment emulating Kodak Vision3 500T, 35mm grain,
halation, anamorphic flare. Mood: gritty 1980s jungle-lab thriller.

Audio: fluorescent hum, water drips, mandible clicks, chitin scrape,
metal splitting, glass shattering, electric crackle, panicked breathing,
low synth + orchestral percussion undercurrent.

The segment-1 payoff: "slow push-in on the researcher, then pan up to reveal the creature." The reveal works because the camera move was scripted, not implied.

Three takeaways from this one. First, motivate your light — "sparking wire acts as the scene's key light" is the single most film-school line in the prompt and it transforms the render. Second, name a film stock. The model knows what Vision3 500T looks like better than it knows what "moody" means. Third, choreograph camera per segment, not globally. One camera instruction for a 15-second sequence is how you get mush.

The mistakes we see constantly

• Prose instead of shots. A beautiful paragraph of story gives the model nothing to cut on. Timestamps or it didn't happen.
• Contradicting your own style block. "Gritty documentary realism" plus "stunning flawless beauty" — pick one.
• No audio block. You're leaving half the model on the table.
• Re-rolling the whole prompt when one shot fails. Fix the failing shot's language. The rest was working; leave it alone.
• Skipping reference images for anything that must stay consistent. Faces, products, wardrobe. Text descriptions drift; references lock.

Camera language cheat-sheet

Words that reliably move the needle, from our internal notes: push-in, pull-back, pan up/down, low angle, tracking shot, follow shot, curved arc, locked-off, handheld weight, whip pan (use sparingly), rack focus, shallow depth of field, motivated lighting, practical light source, 24fps with natural motion blur, anamorphic, macro.

Words that do nothing: epic, cinematic (alone), high quality, masterpiece, 8K ultra HD best quality trending on artstation. The 2023 incantations are dead, let them rest.

Camera language earning its keep: "gentle diffused lighting, slight handheld zoom-in aesthetic that mimics casual fan recording" — two phrases, and the whole clip reads as found footage.

All the gallery prompts — including the K-pop fansign we only showed a frame of, and the anamorphic creature chase we didn't cover — are on the homepage, full text, one click to copy. Steal a structure, swap in your subject, and spend today's free video finding out what needs fixing. That loop, prompt → render → fix one block → re-roll, is the entire craft. Everything else is taste.