Blog

Guide

Kling 3.0 Motion Control Prompt Recipes: Copy-Paste Prompts, Reference Video Tips, and Negative Lists

A practical recipe book for driving motion with reference videos while keeping characters consistent

Nov 12, 2026 12 min read
Flat geometric style blog cover image: on a deep pine green background, several large prisms are cut into sage green and mist green color planes, with crisp edges.
On this page

Why Motion Control Needs a Different Prompt Structure

A motion-control job asks the model to do two things at once: render a specific subject and follow the motion of a reference video. A single descriptive paragraph collapses those two jobs into one blob of text, and when the output drifts you cannot tell which part of the prompt failed.

Treat the prompt as a stack of layers instead:

  • Subject lock — what is in frame and what must stay stable (character traits, wardrobe, props, object identity).
  • Reference motion — how the subject should move, expressed as observable actions and timing that match the reference clip.
  • Style and quality — framing, lighting, rendering intent, and the constraints that keep artifacts out.

Why layers beat paragraphs:

  1. Attribution. When a shot looks wrong, you can point at the layer that caused it. Character face drifting? Subject lock. Action too fast? Reference motion. Muddy texture? Style.
  2. Reuse. The subject lock for a recurring character is the same across clips; only the motion layer changes. You stop rewriting from scratch.
  3. Diffing. Layered prompts are easy to version. You can compare two runs and see exactly which line you changed.
  4. Ordering effects. The model reads your prompt roughly in the order you write it, so putting the identity you need preserved first gives it priority over style flourishes.

The rest of this guide builds around that stack: a reusable base template, a character-consistency discipline, motion and pacing phrasing, duration and aspect-ratio handling, a maintained negative list, and a test loop that changes one variable per run.

The Base Recipe: Three Blocks You Reuse

Here is the template to copy. Keep the block order fixed so every prompt in your project has the same shape.

[SUBJECT]
<character/object lock: stable traits, wardrobe, props, count of people in frame>

[REFERENCE MOTION]
<what follows the reference video: action sequence, timing cues, rhythm, camera behavior>

[STYLE / QUALITY]
<shot framing, lens and lighting intent, texture and rendering constraints>

[NEGATIVE]
<comma-separated failure modes from your maintained list>

A filled example:

[SUBJECT]
one woman, early 30s, shoulder-length dark hair, olive jacket, canvas tote on right shoulder, same face and outfit across all shots

[REFERENCE MOTION]
follow the reference clip: she turns from the window toward the camera in the first beat, walks two steps forward in the middle, pauses and looks down at the tote in the last beat; steady rhythm, no rush, camera stays mostly static with a slight handheld sway

[STYLE / QUALITY]
medium shot, natural window light from camera left, soft shadows, clean fabric detail, realistic motion blur only on the walking steps

[NEGATIVE]
face morphing, extra fingers, outfit color shift, warped tote strap, jittery frame edges, text overlays

Assembly order matters for three reasons:

  • Subject first because it is the thing you least want reinterpreted. Identity cues at the top get more weight than anything that follows.
  • Motion second so the reference video's behavior is described before style language starts pulling the render toward a look.
  • Style last because it is the most negotiable layer. If the clip feels flat, edit style; do not touch the lock.

Two habits keep the base recipe healthy. First, write each block as short declarative fragments rather than prose — fragments survive copy-paste and are easier to diff. Second, keep placeholder text (like <character lock>) in a separate scratch file so production prompts never ship with unfilled brackets.

Keeping Characters Consistent Across Shots

Character drift between generations almost always comes from re-describing the character loosely on each run. Fix it by treating the subject block as a constant.

Write the lock once, then freeze it

Draft the subject block as a fixed list of traits, each phrased the same way every time:

[SUBJECT]
one man, late 20s, short black hair with a slight side part, thin silver-rimmed glasses, charcoal crew-neck sweater, small scar above left eyebrow

Store this exact string. Do not paraphrase it between shots — "silver glasses" in shot one and "metal eyewear" in shot two are different signals to the model.

Choose traits that survive motion

Not every detail is worth locking. Prefer traits that stay visible through typical motion: hair shape, eyewear, garment color and cut, and one distinctive mark. Skip transient details like hand position or expression, which the reference motion will control anyway.

Separate what the motion owns from what the lock owns

The subject block should not mention actions, and the motion block should not re-describe the character. If the motion block says "she raises her hand," the model should infer her hand from the lock. Mixing them invites the model to reinterpret the character while parsing the action.

What causes drift

  • Synonym churn. Re-wording the same trait across prompts.
  • Over-specification. Ten traits of equal weight; the model averages them and loses the distinctive one.
  • Style collisions. A style block that pushes a strong look (heavy stylization, strong color grade) that overwrites locked traits.
  • Multiple people. If the reference clip has two subjects, say explicitly which one the lock describes and how many people should appear in frame.

A quick check before each run: read the subject block alone. If a stranger could not picture a specific, single character from it, it is too loose to hold across shots.

Describing Motion and Pacing in Words

The reference video carries motion, but your wording decides how that motion is interpreted and paced. Vague adjectives give the model room to guess; observable actions with timing cues do not.

Replace adjectives with actions

  • Weak: she moves energetically

  • Better: she steps forward twice, quick and light, then stops with weight on the front foot

  • Weak: dynamic camera

  • Better: camera pushes in slowly during the turn, then holds still

An adjective describes a feeling; an action describes something a viewer could count.

Anchor timing to beats

Break the clip into two or three beats and describe what happens in each. Use ordinal language instead of invented timestamps:

[REFERENCE MOTION]
first beat: subject turns head from left to right
middle beat: shoulders follow, one step back
last beat: settles, breathing visible, gaze returns to camera

Ordinals keep the prompt stable when you change clip duration, and they map cleanly onto whatever the reference video does.

Describe rhythm, not speed numbers

Where the reference has a clear cadence, name it: steady and unhurried, two quick steps then a pause, continuous smooth glide. Pair the rhythm with the action so pace is attached to something observable.

Constrain the camera explicitly

Motion control often fails because the camera moves when it should not. Add one line stating camera behavior — camera static, slight handheld sway, slow push-in only during the last beat — and keep it consistent with what the reference clip actually shows.

Use transition words sparingly

Words like then, after, while, and finally are your sequencing tools. One per beat is enough; stacking them makes the motion read as a list of unrelated events rather than one continuous performance.

Duration and Aspect Ratio in the Prompt

Duration and framing change what a prompt means. A motion block written for a short clip becomes cramped when the clip gets longer, and a subject block written for a wide shot loses detail in a vertical crop.

Write for the clip length you request

Match the number of beats to the clip length you are generating. A short clip supports roughly one action arc: start state, one change, settle. Longer clips can carry two or three beats. If you keep a three-beat prompt but request a short clip, expect the model to compress or drop the final beat.

A practical pattern:

  • Short clip: one beat plus a settle.
  • Medium clip: two beats plus a settle.
  • Longer clip: three beats with an explicit rhythm line between them.

State framing expectations in the style block

Aspect ratio affects composition, so write the shot type into the style block instead of assuming it:

[STYLE / QUALITY]
vertical framing, subject centered with headroom, hands visible in lower third, background softly out of focus

For wide framing, say what fills the extra horizontal space — environment, a second object, negative space — or the model may stretch the subject to fill it.

Adapt when either variable changes

When you change duration or ratio, edit one block only:

  • Longer clip: add a beat to the motion block; do not touch the subject lock.
  • Vertical crop: move the key action into the vertical safe area in the motion block, and adjust framing language in the style block.
  • Wider frame: add environment detail to the style block rather than adding more characters to the subject block.

Record the requested duration and ratio next to each saved prompt so you know which version of a recipe a given output came from.

Building a Negative List That Actually Helps

A negative list is a maintenance artifact, not a one-time paragraph. The useful ones are derived from failures you actually observed.

Start from observed output, not imagination

Run your base recipe once. Watch the output and write down concrete flaws: face morphing on the turn, tote strap warped, fabric color shifts warmer mid-clip, frame edges jitter. Translate each into a short negative term:

[NEGATIVE]
face morphing, warped strap, color shift, jittery edges

Generic lists copied from elsewhere tend to include terms that do nothing for your footage while crowding out the ones that matter.

Keep entries short and specific

Prefer two-word failure descriptions over sentences. extra fingers beats hands sometimes have the wrong number of fingers in close shots. Short entries are easier to reorder, remove, and compare.

Version the list alongside the prompt

Keep the negative list in the same file as the prompt blocks, with a date or revision marker. When a new failure appears, add a term and note which prompt revision it belongs to. When an old term stops helping, remove it rather than letting the list grow forever.

Prune deliberately

Long negative lists can suppress legitimate detail. If your subject lock calls for a textured jacket and your list bans texture, you are fighting yourself. Review the list whenever the style block changes.

Keep project-level and shot-level lists separate

A project list holds failure modes that recur across every clip (identity drift, edge artifacts). A shot list holds one-offs (a specific prop warping). Merge them at prompt time, project list first.

Recipe Variations and How to Test Them

The base recipe is a starting point. Different reference videos need different motion language, but the way you test variations should stay disciplined.

Variations worth trying

  • Action-dense reference: trim the motion block to the two or three beats that matter and let the reference carry the rest.
  • Subtle reference (small gestures, breathing): add explicit rhythm words and a camera-stability line; subtle motion is the easiest to lose.
  • Multi-subject reference: extend the subject block with one line per person and state who the motion follows.
  • Prop-heavy reference: add a short prop clause to the subject lock naming the object and how it should behave.

Change one variable per run

When comparing outputs, edit a single line — one motion phrase, one style term, one negative entry. Two simultaneous edits make it impossible to attribute the difference. Log the change next to the output so the comparison survives past the session.

Keep a small control set

Maintain two or three reference videos you reuse across recipes. Running every variation against the same controls tells you whether a prompt change is real or an artifact of a particular clip.

Score outputs against the intent, not vibes

For each run, check three things: did the locked subject survive, did the motion match the reference beats, and did the negative list hold. Three pass/fail marks per run are enough to see patterns after a handful of iterations.

Integrating Recipes Into a Workflow

A recipe only pays off when it is repeatable. That means treating prompts, reference videos, and negative lists as versioned assets and driving the model through one consistent interface.

Store assets together

Keep a project folder per character or campaign containing:

  • the frozen subject lock;
  • one or more reference videos;
  • the motion block variants;
  • the project negative list with revision notes;
  • a log of runs with duration, aspect ratio, and the pass/fail marks.

If a shot needs to be recreated months later, everything needed is in one place.

Call the model through a single API surface

Automating runs means sending the same structured prompt, reference video, duration, and ratio every time. Using HiAPI.ai as the unified interface keeps prompt, reference asset, and parameters in one request shape across models, so your recipe files stay portable and your scripts do not need rewiring when the stack changes. The Kling 3.0 motion control model page is the place to start wiring requests.

Keep prompts as data

Store prompt blocks as plain text or templated strings in your repo rather than pasting them into a UI. That way a change to the character lock propagates to every shot, and diffs show exactly what moved between generations.

Close the loop

Route each run's output back into the feedback loop: note failures, update the negative list, bump the prompt revision. The recipes in this guide are scaffolding — the loop is what makes them reliably yours.

Related articles

View all articles →

Start building with persistent AI artifacts

Get your API key and call image, video and audio models through one task-based API.

Read the API docs