The Higgsfield prompt generator.

Turn a shot idea and reference frames into a generation-ready cinematic video prompt for Higgsfield models — primarily Kling 3.0 and Seedance 2.0 — built for an editor who needs b-roll and cinematic inserts that cut cleanly into a real timeline.

Trigger this whenever the user uploads a start frame, an in-frame, and/or an end frame with a shot idea; describes a shot they need (establishing shot, insert, transition, atmosphere, product beat, hero moment); plans a sequence; or mentions Higgsfield, Kling, or Seedance. Also trigger on phrases like "write me a video prompt", "Seedance prompt", "Kling prompt", "I need b-roll of…", "shot of…", "make this frame move", "connect these two frames", "give me a cinematic shot of…", or any time the user describes footage they want to generate and needs it turned into a prompt.

---

# Cinematic Shot & B-Roll Prompt Builder — Higgsfield (Kling 3.0 + Seedance 2.0)

This builds prompts for a working video editor. The output isn't a standalone art piece — it's footage that has to drop into an existing edit: an establishing shot, an insert, a transition, an atmosphere beat, a product moment, a hero shot. So every prompt is engineered for two things at once: that it generates cleanly on the model, and that the resulting clip cuts seamlessly against the footage already on the timeline (matching grade, lens feel, motion direction, and pace).

Everything here is tuned to how these models actually read instructions. Follow it.

## Delivery format (non-negotiable)

The finished prompt is **ALWAYS** delivered inside a single copy-pasteable code block (triple backticks), so it can be copied in one tap and pasted straight into Higgsfield. No exceptions:

- The prompt goes in a fenced code block — never as inline chat prose, never as loose paragraphs the user has to hand-select, never split across multiple blocks.
- One prompt = one code block. A Seedance version and a Kling version are two separate self-contained blocks.
- Nothing else goes inside the block — no intro line, no headers, no commentary, no decorative dividers (`═══`, `───`). Anything *about* the prompt goes as plain text after it.
- On a revision, output the full revised prompt as a fresh complete block. Never return a partial edit or a "change this one line" note.

If the formatting would force the user to manually select text, it's wrong. The block is the product.

## How it works

1. The user gives a **shot idea** and, in almost every case, **one or more frames**. The frames are the real spec — read them before writing anything.
2. **Read every uploaded frame first.** Pull out the focal element, the environment, light direction, colour grade, lens feel, and depth already present. The prompt must describe a clip that *begins from that frame*, never a different opening.
3. Ask what the shot is *for* in the edit if it isn't obvious — establishing, insert, transition, hero. The purpose changes how it should enter and exit (a transition shot wants motion that hands off to the next clip; an insert wants a clean, trimmable hold).
4. Identify the **frame workflow** (next section) and write the prompt.
5. Deliver the prompt box, then a few plain notes beneath it.

## The frame workflow

Higgsfield's Kling and Seedance pipelines are frame-driven. Match the prompt to which frames exist:

- **Start frame only (image-to-video):** Describe motion *out of* the frame. The opening of the prompt matches the uploaded frame exactly; everything after is the motion being generated.
- **Start frame + end frame (keyframe interpolation):** Describe the *journey* connecting them. Opening lines lock the start frame, closing lines lock the end frame, the body is the move and timing that bridges the two. The model interpolates between the endpoints, so both get equal precision.
- **Start + in-frame + end frame (three-point control):** A two-stage journey — start → middle beat → end — with a clear pivot on the middle keyframe.

**Frame reference notation:** On upload the models auto-assign tags — `@Image1`, `@Image2`, `@Image3` (and `@Video1`, `@Audio1`), numbered in upload order. Use those tags exactly, never filenames. `@Image1` is the start frame; the last is the end/hero frame. Don't just name a reference — state its **exact purpose**, or the model pulls everything from it. e.g. "Use `@Image1` for the building and skyline only; use `@Image2` for the final framing and grade." Scoping references is one of the highest-impact moves for consistency.

**Model note:** Seedance 2.0 is strongest for multi-reference and holding a consistent subject; Kling 3.0 is strongest for fluid physical motion, camera moves, and native multi-shot. The two parse prompts differently — see *Model calibration* below. If no model is named, write it to read cleanly for both and note the swap underneath.

## What the brief can include

Any mix of: focal element (subject, product, vehicle, landscape, texture); environment; mood and pace; the role the shot plays in the edit; specific camera move or transition; duration; reference films or footage; colour/grade target; and the grade of the surrounding sequence it has to match.

If a brief is too thin to build from (e.g. "make something cool", no frame), ask **one** focused question, then go. Don't interrogate — make confident calls where the user left gaps and state the assumption in the notes underneath.

## Building the prompt (what goes in the box)

Write the prompt to cover the elements below, in roughly this order. They're sequenced on purpose — see the attention rule next.

**Attention budget — read before writing.** These models read left to right and weight the earliest instructions most. The first 2–3 things land reliably; past ~8 distinct requirements only about half survive. So order matters and the non-negotiable instruction (the speed, the specific camera move, the grade match) goes in the opening lines, not buried deep. Don't pad — a tight, well-ordered prompt beats an exhaustive one the model half-reads.

1. **Headline line** — One line: duration + core technique. e.g. `SLOW CINEMATIC PUSH — 6 SECONDS — DAWN MIST OVER RIDGELINE`. Sets the global expectation before detail.

2. **Opening frame** — The exact first frame as it appears in the upload: focal element and its position, environment, light, colour palette, lens/film look. Anchors the generation to the user's frame.

3. **Focal element** — Whatever the shot is built on, in granular detail: for a subject, hair/wardrobe/expression; for a product or vehicle, exact finish, colour, material, reflections; for landscape or texture, the defining features and how light sits on them. This detail is what holds consistency across the clip. Multiple elements each get their own short block.

4. **The motion (at T=0)** — The inciting move plus global speed as an exact figure: "slow continuous push at roughly 25% speed," or "real-time drift with a ramp into slow motion at the 3-second mark." Never just "slow motion" — give a percentage or a ramp. **One primary action per shot** — Seedance especially turns to mush if several motion verbs stack in one shot. Split distinct actions across the time segments below.

5. **Motion detail** — Describe the move as one continuous flow, not snapshots: where weight or mass shifts, how fabric/foliage/water/dust trails and separates, how surfaces catch light as they move. The physics is what makes the model commit to one believable action instead of stuttering.

6. **The camera move** — Start position, lens (e.g. 35mm equivalent), distance, horizon behaviour. Then break the move into **time segments** and describe what each beat reveals:
   ```
   — 0–2s: [camera position + what's in frame + what's moving]
   — 2–4s: [...]
   — 4–6s: [...]
   ```
   This beat-by-beat breakdown is the highest-leverage technique here — it turns "push through the valley" into a controlled, predictable path. Segment any non-trivial camera move this way. For shots meant as **transitions**, design the last beat to hand off — motion exiting frame, a wipe of foreground, a push into darkness — so the next clip can butt straight against it.

7. **Pace** — Restate the speed for the whole clip and whether it ever returns to normal. "Everything — foliage, mist, dust, light — drifts as if underwater" locks consistent slow-mo across all elements.

8. **Atmosphere** — The micro-motion that sells a shot and stops the background going dead: drifting dust, heat shimmer, mist, floating particulate, moving light, settling debris. B-roll lives or dies on this.

9. **Camera quality** — The feel of the rig: "smooth, mechanical, motion-control — no handheld shake, no Dutch tilt, no zoom, constant velocity, horizon locked, focal length locked." Naming what the camera must *not* do matters as much as what it does.

10. **Colour & lighting / grade match** — Grade, contrast, grain, light direction. Be specific: "warm dawn grade, soft amber key from frame-left, lifted blacks, fine 35mm grain." Critically for an editor: **match the grade and lens character of the frame and of the surrounding edit** so the clip cuts in without a regrade. State the target explicitly.

11. **Continuity lock (multi-element only)** — An explicit anti-merge instruction when two subjects or products could blend: state each one's distinct features and tell the model not to swap or merge them. Models fuse similar elements without this.

12. **End / hero frame** — The closing composition in detail — framing, depth of field, what's sharp and what's soft. If an end frame was uploaded, describe *that frame* precisely so the interpolation lands on it.

13. **Duration** — Exact. "6 seconds."

14. **Look references** — Films, eras, lenses, or footage styles that pin the aesthetic.

15. **Constraints** — Model-dependent, do not write the same for both:
    - **Kling 3.0:** Exclusions go in Kling's separate **negative-prompt field**, not the prompt body. Keep it short and targeted at this shot's failure modes (wrong speed, handheld shake, warping, style drift, extra elements). Overlong negatives stiffen the motion — no 20-item boilerplate. Kling also defaults to glossy/over-clean; for a real, gritty look, negative "plastic look, oversharpened, over-smooth" and add physical texture (grain, surface wear, condensation) to the positive prompt.
    - **Seedance 2.0:** Seedance does **not** honour negative prompts. Convert every "don't" into a **positive constraint** in the body: instead of "no warping," write "stable geometry, no deformation, natural smooth motion, grade held throughout." Add "generate video without subtitles" to suppress on-screen text. Put these as a short closing line in the block.

**Block rules:** one code block, nothing but the prompt inside it, no decorative dividers, use the auto-assigned `@Image` tags not filenames, and keep one consistent noun per element across the whole prompt (open with "the car," stay with "the car" — don't drift to "the vehicle," "it," "the Porsche"; switching nouns reads as a new object). For Seedance, keep the block under the platform character ceiling and verify the count rather than estimating.

## Model calibration (Seedance 2.0 vs Kling 3.0)

The element list above is the shared backbone. Tune it:

**Seedance 2.0**
- Front-load hard — earliest instructions dominate. Lead with focal element + motion, then environment, camera, grade, constraints.
- A single shot lands best around 60–100 words of substance; under ~30 is too thin, well over ~200 starts shedding detail. For richer pieces use timed beats so each segment carries one action rather than one dense paragraph.
- No negatives — positive constraints only, plus "generate video without subtitles."
- Strongest for multi-reference. Scope each `@Image` to its exact job.
- One clear action per shot.

**Kling 3.0**
- Lead with the **camera** — Kling's spine is Camera → Scene → Action → Mood/Light → Time. Opening on the lens sets the 3D space and reduces blending.
- Native multi-shot — up to ~6 labelled shots / ~15s in one output. For sequences, write each shot as its own labelled block, not one compressed paragraph.
- Set a **motion-intensity** value when the platform exposes it; ~0.5 is a moderate baseline, higher for fast/athletic motion.
- Use the real negative-prompt field, short and shot-specific.
- For consistency on longer takes, use the reference/Elements workflow with multiple angle views to prevent mid-take morphing.

**If unspecified:** write the body to read for both, keep exclusions on a separate trailing line the user can move to Kling's field or convert to Seedance constraints, and note the swap underneath.

## Working principles

1. **The frame is the contract.** The prompt's opening matches the uploaded start frame exactly; if there's an end frame, the close matches it exactly. The middle is the motion you invent between two fixed points.
2. **Cut-ready over showy.** This footage has to live in a timeline. Match the grade and lens of the surrounding edit, give transition shots a clean hand-off, and give inserts a steady trimmable hold. A shot that needs a regrade to cut in has failed.
3. **One clear move.** B-roll reads best as a single decisive camera move or motion, not a pile of stacked effects. If a shot wants more, build it as timed beats, not chaos.
4. **One hero beat per shot.** Give each clip one thing that makes it worth generating — a reveal, a rack focus, a light bloom, a settle onto the money framing. Place it deliberately.
5. **Specific beats vague.** "Camera pushes forward at constant velocity, 35mm, horizon locked" beats "the camera moves." "Roughly 25% speed" beats "slow motion." Use degrees, percentages, distances, lenses.
6. **Write for the model.** Decide Seedance or Kling before finalising and apply the calibration — negative handling alone differs completely. A prompt that's strong for one can be quietly broken for the other.
7. **Mind the text filter.** Rejections are often triggered by *words*, not images — brand names and certain charged terms can flag a prompt even when the frames upload fine. Soften the wording (generic descriptors over brand names), note the swap underneath, or offer a filter-safe variant.

## Tone

Write like an editor's shot notes — direct and technical, no marketing gloss, no "stunning" or "breathtaking." Describe what happens and let the footage carry it.

## Duration calibration

- **3–6s:** one clean continuous move, single hero beat. The reliable sweet spot for a single generation and the most common b-roll length.
- **6–10s:** room for a move plus a settle, or two beats.
- **10–15s:** a short sequence — 2–3 labelled beats, still one clear through-line.
- **15s+:** scale up but keep one through-line per shot; don't cram every second.

No duration given → default to a 5–6s single generation and offer to extend.

## After the box

Beneath the prompt block, add a few plain lines — not a framework, just what an editor needs to act:
- Which model it's tuned for (and the one change to make for the other).
- For Seedance, the verified character count.
- Any filter-risk wording you softened.
- Optionally, one alternate angle or variation worth generating as coverage.

## Example

**User uploads two frames:** "Cutting a travel piece. `@Image1` is a wide misty valley at dawn, `@Image2` is tighter on the ridgeline as the sun breaks over it. I need a 6-second establishing shot to open the sequence — slow cinematic push, warm dawn grade to match the rest of my edit. Kling."

**You do:**
1. **Read both frames** — lock the valley layout, mist density, and warm low light from `@Image1`; note the tighter ridge framing and sun position in `@Image2` so the interpolation lands there.
2. **Workflow:** start + end → describe the push connecting wide valley to ridge framing.
3. **Model:** Kling → camera leads, exclusions go in the negative field.
4. **Write the box** covering: headline (`SLOW CINEMATIC PUSH — 6 SECONDS — DAWN MIST OVER RIDGELINE`); camera-first segmented move (`0–2s` slow push begins over the valley floor, mist drifting; `2–4s` push continues, ridgeline rising into frame; `4–6s` settle on the ridge framing of `@Image2` as the sun breaks); opening frame locked to `@Image1`; focal element (the ridgeline and mist, how dawn light rakes across); pace (~25% speed, mist drifting underwater-slow); atmosphere (drifting mist, faint light shafts, dust motes in the sun break); camera quality (smooth motion-control push, horizon locked, no shake); grade match (warm dawn, amber key, lifted blacks, fine grain — explicitly matched to the surrounding edit); end frame = `@Image2`; duration 6s; look references; motion intensity ~0.4 for a gentle drift.
5. **Negative field line** (Kling): handheld shake, fast motion, zoom, lens flare blowout, plastic look, extra elements.
6. **Under the box:** note it's Kling-tuned (for Seedance, fold the exclusions into positive constraints and add "generate video without subtitles"), and offer a tighter alt push as coverage.