# Kling Multi-Shot Prompting Guide Status: working prompt-SOT for production experimentation Updated: 2026-03-31 Scope: Kling Series 3 / Omni multi-shot prompt writing for cast control, shot differentiation, motion design, and ending stability ## Purpose This document turns recent multi-shot experiments into reusable prompt-writing guidance. It is **not** raw API SOT for field names. That remains in: - `kling-api-field-level-tier1.md` - `kling-qingque-full-api-reference.md` This guide is the **operational SOT for prompt design**: - how to write prompts that better preserve cast identity - how to avoid extra people / duplicate people - how to improve shot differentiation - how to reduce freeze-like endings - how to trade off `std` vs `pro` --- # 1. What we learned from live runs ## 1.1 Cast lock can improve, but motion usually suffers When prompts strongly emphasize: - exactly N people only - no extra people - no duplicates - keep the same referenced women across all shots - preserve left-to-right order we observed: - fewer extra people - fewer duplicate people - better identity stability but also: - more static group formations - more repetitive shot grammar - weaker motion energy - more conservative scene progression Interpretation: > stronger cast control often pushes the model toward safer, more static layouts. Additional operational reading from the 2026-03-31 runs: - `<<>>` plus explicit cast-lock wording materially helps compared with a generic group description - `pro` can improve visual polish, but it does not automatically solve cast-control failures by itself ## 1.2 Motion wording alone is not enough Adding verbs like: - stride - walk - pivot - turn - subtle motion helps somewhat, but if all shots stay in the same motion family, the model may still produce: - zoom-in variations - angle variations of the same event - repeated entrance/lineup imagery Interpretation: > motion verbs are useful, but **shot-level event differentiation** matters more. ## 1.3 Ending shots are a freeze-risk zone Prompts that combine ideas like: - final - hero - lineup - locked frontal shot - stopped walking - subtle motion only frequently encourage: - freeze-like holds - near-static final seconds - angle changes without meaningful motion Interpretation: > the final shot must be written as **continuing motion**, not just a finished pose. ## 1.4 `pro` improves quality, not necessarily control Observed trend: - `pro` often improves image quality / polish - `pro` does **not** automatically fix identity mixing, extra people, or weak storyboard fulfilment Interpretation: > `pro` can raise visual quality, but prompt structure still determines whether shots are actually distinct and controlled. --- # 2. Hard prompt-design principles ## 2.1 Separate the goals in your head Do not assume one prompt can maximize all of these at once: - cast fidelity - motion energy - shot differentiation - premium commercial polish These goals often trade off against each other. Operational note: > richer background / luxury-scene prompting can improve set richness, but may also increase pressure on cast purity in already-complex 4-person prompts. Recommended prompt-design order: 1. decide whether cast fidelity or motion is more important 2. lock the minimum continuity needed 3. add only enough motion/event complexity to reach the creative goal 4. explicitly protect the ending from freezing ## 2.2 Each shot must differ in 3 ways For multi-shot prompts, every shot should differ along **all three axes**: 1. **event** - what is happening? 2. **camera grammar** - what kind of shot is it? 3. **end state** - where do the subjects end up before the next shot? If you only change one axis, Kling often collapses into a variation of the same scene. Bad pattern: - Shot 1: walk forward - Shot 2: walk forward from a new angle - Shot 3: walk forward from another angle - Shot 4: final lineup Better pattern: - Shot 1: enter in formation - Shot 2: cross / layer / re-block - Shot 3: stop and pivot-turn together - Shot 4: arrive into final moving hero finish ## 2.3 The final shot should be an arriving state, not a static state Prefer: - continuing to step into final marks - small synchronized steps still visible - gentle posture adjustment while moving - visible motion throughout the whole shot Avoid writing the ending as: - fully stopped - locked - subtle motion only - final lineup hold unless a static hold is actually desired. --- # 3. Cast-control guidance ## 3.1 Recommended cast-lock wording When identity stability matters, these prompt patterns are useful: - exactly four women only - no extra people - no duplicates - keep the same four women across all shots - preserve the same identities - no identity mixing - all four women remain visible - do not lose any of the four referenced women - do not replace one woman with a duplicate of another These help reduce: - extra generated people - duplicate faces - cast drift - temporary cast disappearance / replacement failures ## 3.2 Position wording helps, but is not hard control Useful phrasing: - `<<>>` on the far left - `<<>>` on the left-center - `<<>>` on the right-center - `<<>>` on the far right This appears to help steering. But current doc/live evidence does **not** support treating this as deterministic slot-locking. Interpretation: > prompt position wording is a steering aid, not a guaranteed cast-layout control system. ## 3.3 Use `<<>>` when explicit reference matters Prompt templating with: - `<<>>` - `<<>>` - `<<>>` - `<<>>` is worth using when you want: - role assignment - left/right identity references - more explicit shot descriptions Operational reading: > element templating can improve clarity and cast control, especially compared with a generic group description. --- # 4. Motion-design guidance ## 4.1 Use action verbs, but diversify the action family Do not rely only on repeated walking verbs. Instead, distribute actions across the sequence: - stride forward - shift formation - pivot turn - rotate shoulders - glance and turn - slow into marks - continue stepping while settling If every shot is just a version of walking, the shots may blur together. ### Additional caution for 4-person group prompts Be careful with subgroup blocking such as: - two women move forward while two remain behind - one side advances more than the other - cross toward center while others stay back In live runs, this kind of wording increased the risk that: - one woman becomes a foreground leader - spacing becomes unbalanced - one referenced woman disappears - another referenced woman is duplicated For stable 4-person commercial group motion, prefer: - all four move together - same speed - balanced spacing - unified formation - no single woman moves significantly ahead of the others Operational note: > subgroup blocking should currently be treated as a higher-risk pattern for 4-person prompts. ## 4.2 Avoid “subtle motion only” unless you really want near-static results This wording is dangerous in late shots. If used, pair it with explicit continuous movement language, e.g.: - subtle but continuous visible motion throughout the whole shot - small synchronized steps remain visible - ongoing posture adjustments and flowing hair through the entire shot Otherwise the model may interpret it as: - almost still - safe hold - frozen-looking commercial ending --- # 5. Camera-grammar guidance ## 5.1 Name the shot type clearly Good camera grammar examples: - wide frontal entrance shot - medium side-tracking shot from the left - diagonal three-quarter angle - frontal tracking shot - locked frontal hero shot Do not only say “camera changes angle.” That is usually too weak. ## 5.2 Give each shot a distinct camera identity Recommended pattern: - Shot 1 = wide frontal entrance - Shot 2 = side-tracking layered motion - Shot 3 = diagonal pivot/turn shot - Shot 4 = frontal finish with continued motion This helps reduce simple zoom-in / zoom-out repetition. --- # 6. Ending-shot anti-freeze rules ## 6.1 High-risk wording combinations These combinations tend to produce freeze-like endings: - final + lineup - stopped walking + subtle motion only - locked shot + hero pose - settle into final lineup + no explicit continuing motion ## 6.2 Safer ending wording Prefer phrases like: - continue stepping slowly into final marks - visible motion continues through the entire final shot - the women keep moving naturally while arriving into the final formation - gentle walking slowdown, not a static hold - no static hold, no frozen pose ## 6.3 Recommended final-shot structure A good final shot often needs all three: 1. clear ending composition 2. explicit ongoing movement 3. explicit anti-freeze instruction Example: > The four women arrive into a final premium formation while continuing small synchronized steps, visible posture adjustments, and flowing hair motion throughout the entire shot, with no static hold and no frozen pose. --- # 7. Recommended multi-shot writing workflow ## Step 1 — Define the sequence goal Choose one dominant goal: - cast lock - energetic motion - shot differentiation - premium finish ## Step 2 — Write top-level continuity guardrails Examples: - exactly the four referenced elements only - no extra people - keep the same four women across all shots ## Step 3 — Design shot-by-shot event changes For each shot, specify: - what happens - what changes versus the prior shot - how the shot ends ## Step 4 — Assign camera grammar Each shot gets a distinct camera identity. ## Step 5 — Protect the final shot Add explicit wording to prevent the final shot from collapsing into a static hold. --- # 8. Do / Don't summary ## Do - use explicit cast count wording - use `<<>>` when role clarity matters - define left/right/group positions when useful - separate each shot by event, camera grammar, and end state - make the final shot continue moving - write shot 4 as an arriving state, not a frozen state - use `pro` when quality matters enough to justify extra cost ## Don’t - assume `pro` solves identity or storyboard problems by itself - write every shot as a variation of walking forward - rely on “camera changes angle” as the main difference between shots - combine final + locked + lineup + stopped walking + subtle motion only unless static hold is intended - assume prompt position wording is deterministic slot control --- # 9. Reusable prompt skeletons ## 9.1 Four-person cast-lock skeleton Top-level: > exactly the four referenced elements only, no extra people, keep the same four women across all shots Shot seed pattern: > `<<>>` far left, `<<>>` left-center, `<<>>` right-center, `<<>>` far right Use this when identity drift is the main risk. ## 9.2 Shot-differentiation skeleton For each shot include: - cast/control sentence - camera grammar sentence - event sentence - end-state sentence - anti-duplication sentence Template: > Keep the same exact four women only. [camera grammar]. [event]. [end state]. No extra people, no duplicates, no identity mixing. ## 9.3 Anti-freeze ending skeleton > Final frontal hero finish with exactly the same four women only. The four women continue small synchronized steps and visible posture adjustments while arriving into the final formation, with flowing hair and fabric motion throughout the entire shot. No static hold, no frozen pose, no extra people, no duplicates. --- # 10. Current operating interpretation ## What is likely to work better - explicit cast count - explicit element references - explicit left/right order - explicit event change per shot - explicit camera grammar per shot - explicit no-freeze ending language ## What remains risky - four-person multi-shot complexity - repeated walking-only structure - long static-ish endings - assuming prompt wording alone guarantees cast correctness --- # 11. Suggested next experiments 1. Compare templated prompts with and without explicit left/right order 2. Compare static hero ending vs anti-freeze moving ending 3. Compare 4-person vs 3-person cast stability under the same storyboard grammar 4. Compare `std` vs `pro` using the same shot/event structure 5. Probe whether shorter final shots reduce freeze risk further --- # 12. Bottom line Current best working interpretation: > Kling multi-shot prompt quality improves when cast control, shot differentiation, and ending anti-freeze design are written as separate layers instead of being mixed vaguely into one commercial description. Put simply: - cast lock prevents chaos - event design prevents repetition - ending design prevents freezing All three are needed.