Higgsfield Prompt

Higgsfield Prompt Engineering

QUICK FACTS

Generated-checked block (scripts/build_index.py verifies anchors). Read the linked sections for full context — these lines are routing aids, not the rules themselves.

• MCSLA = Model, Camera, Subject, Look, Action — the five layers of every prompt →
• I2V: describe ONLY what moves or changes, never what's already in the image →
• Keep prompts under 200 words — short-form MCSLA regime only; block-scaffold production prompts replace the cap with structural lint (HARD RULE 8 carve-out); Cinema Studio has a hard 512-character cap →
• 1 primary action per clip, 1–2 secondary max; Fast Motion Trick: render in Slow Mo, speed up in post →
• Never leave a generic emotion ("sad"/"angry") in a prompt — decompose into muscle movements, breath, eyes, skin →
• Soul ID / recurring characters: split every prompt into Identity Block + Motion Block — never mix them →
• Conflict order when sub-skills disagree: explicit user direction > scene archetype > emotion-sync →
• Aspect ratio is a per-model enum set in the UI/header, never in the prompt body — verify via ../../specs/model-specs.yaml →
• Never combine Dolly In + Dolly Out in one shot; @ Elements for static scenes, plain text for action →
• Iterate by changing exactly ONE variable per regeneration →
• 6-Pass Diagnostic order: Subject → Action → Camera → Style → Audio → Output; most failures land on Pass 1–2 →
• Seedance short-form: 30–100 words win; Subject + Action in the first 20–30 words. Block-scaffold production briefs run 218–2,059-word medians by register — see ../higgsfield-seedance/SKILL.md § Official Prompt Architecture →
• Genre length targets: Product 30–50w, Lifestyle 40–60w, Drama 60–100w, Music Video 50–80w, Anime 50–90w →
• Kill slop words (beautiful, stunning, epic, amazing) — replace with concrete visuals/physics →
• Seedance/CS 3.0 has NO negative-prompt syntax — phrase as positive constraints →
• Dialogue cap: ~25–30 spoken words fit in 15 seconds — keep the power-shift line, convert the rest to behavior →
• Engine limits: ≤3 characters tracked across cuts; exit-frame = gone; off-screen = nonexistent; avoid reflections →
• Every cut must change BOTH shot size AND camera character →
• Age-blind rule: never boy/girl/child/kid/young/teen/little — describe by role, clothing, action →
• Scenes start already in progress unless the user says "starts with…" or "ends with…" →

The MCSLA Formula

Every high-performing Higgsfield prompt is built on five layers. Think of it as the cinematographer's checklist — fill in each layer and the model has everything it needs.

Prompt Types

Text-to-Video (T2V)

Start from nothing — describe the entire scene from scratch. Best for: establishing scenes, abstract concepts, environments without a specific character.

[Subject + appearance].
[Environment — location, time, weather, atmosphere].
[Action — what happens and how].
[Camera — named control].
[Look — style + color grade].

Example:

A lone astronaut stands on the surface of a red desert planet, helmet visor reflecting
twin moons rising on the horizon. Dust spirals slowly in the thin atmosphere.
She turns to face the camera, gloved hand raised in a slow salute.
Camera: slow Crane Up revealing the vast emptiness behind her.
Style: Cinematic, desaturated orange and deep blue, 2.35:1 anamorphic.

Image-to-Video (I2V)

Animate a provided still image. The image defines the starting frame. Best for: character consistency, product shots, portrait animation, storyboard bring-to-life.

[Reference the input image as the first frame].
[Describe what should move, change, or animate — not what is already visible].
[Camera — named control].
[Style/atmosphere cues].

Example:

Starting from the provided image as the first frame.
The woman's hair lifts gently in the wind. She blinks slowly and turns her gaze
slightly to the left, a faint smile forming.
Camera: subtle Dolly In toward her face.
Style: Cinematic, warm afternoon light, shallow depth of field.

Key rule for I2V: Do NOT re-describe what is already in the image. Only describe what should change or animate. Over-describing the static elements confuses the model. This applies equally to an @Image used as the starting frame in Seedance/Cinema Studio 3.0 — describe ONLY motion and camera movement, never what's already visible. An @Image attached as an identity reference (not frame one) is a different regime: never write text that contradicts it; how much matching identity text may ride beside it is OPEN (../shared/house-rulings.md P1-1).

Narrative Structure

Fluid Narrative (preferred for most use cases)

Write the scene as continuous action. No timestamps. Most natural for Higgsfield.

A detective pushes open the door to the rain-soaked rooftop, coat whipping in the wind.
She steps to the edge and looks down at the city below — a thousand lights blurring
through the downpour. Camera dollies slowly behind her, then cranes up to reveal the
skyline. Cinematic style, cold blue tones, 16:9.

Timestamped (use only for precise multi-beat sequences)

Only use when exact timing of separate actions matters — e.g., a transformation, a multi-phase action sequence, or a beat-synced music video. Maps to Cinema Studio 3.0's Custom multi-shot mode.

0–3s: Wide establishing shot. The fighter stands alone in the ring, chest heaving.
3–6s: Crash Zoom In on his face. Sweat on his brow, jaw clenched.
6–10s: 360 Orbit as he raises his fists. Crowd noise rises.

High-Performing Prompt Patterns

The #1 mistake in video prompting: over-describing appearance and under-describing behavior. Give your subject something to DO. Give them an internal state that creates visible behavior. A verb that describes motion or intention is more important than adjectives.

Specificity beats generality:

• ❌ "the camera moves dramatically"
• ✅ "camera Dolly Zoom In — subject stays the same size as the background rushes forward"

Active verbs carry the scene:

• ❌ "a woman is in an alley"
• ✅ "a woman darts through a rain-soaked alley, coat flapping, boots splashing"

Name the camera control: Higgsfield understands its own preset names. Always use them explicitly.

• ❌ "the camera slowly circles"
• ✅ "360 Orbit around the subject"

Lead with subject, end with style: Subject → Action → Camera → Style is the most reliable order.

Keep it under 200 words (short-form regime): Focused prompts outperform exhaustive ones. One clear intention > ten vague details. Regime exception (HARD RULE 8): block-scaffold production prompts — ../higgsfield-seedance/SKILL.md § Official Prompt Architecture — replace the word cap with structural lint; harvested production briefs run 218–2,059-word medians by register. The cap governs single-shot MCSLA prompts only.

Cinema Studio: Keep it under 512 characters: Cinema Studio has a hard 512-character limit on prompts (both 2.5 and 3.0).

• 2.5: @ Element chips consume ~80–100 hidden characters each. With 2 @ tags, keep visible text under ~250 chars.
• 3.0: @ references (images/video/audio) are media attachments, not inline metadata — they consume less hidden space. Keep visible text under ~350–400 chars with references, ~450–500 without. See the Cinema Studio skill for full character budget details.

The Pre-Prompt Checklist

Before writing any prompt, answer these five questions. Vague prompts like "give me something cinematic" tell the AI nothing.

One Action Per Scene

AI models can replicate real-life physics — but only so much at once. Asking for multiple complex actions in one clip overwhelms the model.

Rule: 1 primary action per clip, with 1–2 secondary actions max.

Break complex sequences into separate shots and stitch them in a video editor, or use Multi-Shot Manual mode to prompt each scene separately.

Fast Motion Trick: If fast motion keeps morphing or breaking, generate the scene in Slow Mo first, then speed it up in post (CapCut, Premiere, DaVinci). The model renders cleaner physics in slow motion.

Generic-Emotion Decomposition — Which kind of X?

Never leave a generic emotion in a prompt. "Sad" / "angry" / "surprised" / "scared" / "thoughtful" / "in love" — each is at least three or four distinct physical realizations, and the model renders a different version depending on which one your prompt invites. A prompt that says only "she looks surprised" produces a different shot every regeneration and degrades adherence across batches.

The rule: decompose the generic emotion into specific muscle movements, breath, eyes, and skin. If you can't decompose confidently, ask the user to choose a variant.

Clarification template — offer when the script or user supplies a generic emotion you cannot decompose without inventing detail:

Which kind of surprise? (a) Light positive — eyebrows lift, lips part softly, slow inhale through the nose, no other movement. (b) Shock — sharp inhale through the mouth, eyes widen, body freezes in place, hand involuntarily lifts to chest. (c) Disbelief — slow blink, head tilts a fraction, lips press together, only one eyebrow lifts. (d) Surprise-with-joy — eye light shifts (catchlight reads), smile builds gradually, shoulders relax.

Same shape applies to any generic adjective — "tense" / "sad" / "angry" / "scared" / "thoughtful" / "in love" each decomposes into 3-5 distinct physical realizations. The decomposed prompt produces a performance; the generic prompt produces AI-video.

Preset library alternative. For named micro-expression presets that drop into a prompt without first-principles decomposition, see ../higgsfield-soul/SKILL.md § Micro-Expressions. The catalog covers most common emotional registers with locked physical descriptors. Use the decompose-from-first-principles rule above when no preset matches; use the preset library when one does.

Layered emotion states

Single-axis decomposition (above) names one register: angry / sad / surprised. Layered emotion names a composite state where two registers stack — anxious determination, tired tenderness, bitter amusement, cornered calculation. Production-team practice finds the model renders layered states better than single registers when the layering is described as one channel modulating another: the dominant state plus the underlying state plus the visible tell.

• Anxious determination — set jaw + locked gaze (determination) with shallow chest breath and a single hand at the side flexing open-closed (anxiety underneath).
• Tired tenderness — soft micro-smile + half-closed eyes (tenderness) with the body weight settled, slow blink interval (tiredness underneath).
• Bitter amusement — one-sided smirk + eye-shine (amusement) with no smile crinkles at the eye corners (bitterness underneath).

Compose layered states by stacking decomposed physical realizations from the single-axis catalog. The dominant state goes in the face; the underlying state goes in breath, posture, and hand-state; the visible tell sits in the eyes.

For finer control, layer a tiny detail on top of an existing emotional cue: Roco is very upset, and his lower lip trembles. The base emotion gets the broad performance; the tiny detail gives the model a specific physical cue to render. Production-team discipline holds that the model renders the simple-emotion-plus- tiny-detail compound better than either an over-decomposed prompt or a too-generic one.

Identity vs. Motion Separation Rule

When a prompt involves Soul ID or any character who must stay consistent across shots, always split the output into two clearly labeled blocks:

Identity Block — Static visual descriptors ONLY

• Face features, skin tone, body type, distinguishing marks
• Clothing, accessories, color palette
• NO motion, NO camera, NO temporal language

Motion Block — Temporal and camera ONLY

• Camera movement, action choreography, speed
• Environmental motion, atmospheric changes
• NO character appearance repetition

Bad (mixed) — identity drifts:

A woman with sharp cheekbones and auburn hair in a blue trench coat runs through
a rain-soaked alley, her coat flapping, sharp cheekbones catching the neon light,
camera chasing her at full speed, her auburn hair streaming behind her.

Good (separated) — identity stays locked:

Identity Block:

The Soul ID character — sharp cheekbones, auburn hair shoulder-length,
wearing a blue trench coat with silver buttons, lean athletic build.

Motion Block:

She runs through a rain-soaked alley, coat flapping behind her.
Camera: Action Run — low behind, matching pace.
Neon reflections streak across wet concrete.
Style: Cinematic, cold blue shadows, warm neon accents. 16:9.

When to apply this rule:

• Always when Soul ID is active
• Always in multi-shot sequences where the same character appears
• Always when camera movement is involved alongside a character
• In Cinema Studio, identity goes in the @ Element definition; motion goes in the prompt

Camera matches emotion, not just identity. The Motion Block describes WHAT the character does and HOW the camera moves. The quality of the camera motion — jittery handheld for anger, smooth handheld breathing for calm, static + slow push for revelation — should track the focal character's emotional state. See ../higgsfield-camera/SKILL.md § Camera-Emotion Sync for the 6-emotion movement map and the emotional-arcs-within-a-shot pattern. For decomposing the underlying generic emotion before picking a camera prescription, see § Generic-Emotion Decomposition above.

Conflict resolution between sub-skills

Sub-skills can legitimately nominate different things for the same shot. higgsfield-camera § Camera-Emotion Sync nominates handheld-slow-low for sadness; higgsfield-prompt § Scene Archetype Router permits locked dolly-in for the Atmosphere archetype where mood-is-the-content. When two sub-skills nominate different camera moves (or motion presets, or style registers) for the same scene, resolve in this order:

1. Explicit user direction wins. If the user said "slow push-in," that's the camera move. The agent's job is to make that direction work with the rest of the structure, not to override it.
2. Scene archetype next. If the user did not specify, pick the archetype-recommended move (Atmosphere → static / slow push-in / locked-off; Action → handheld / whip-pan / FPV; Dialogue → shoulder-coverage / push-in on emotion).
3. Emotion-sync register last. Camera-Emotion Sync nominations are the default tiebreaker when archetype is unclear.

When the resolution is non-obvious, surface it. Tell the user which sub-skill nominated what and why you picked one over the other — this is meta-correct behavior and lets the user override. Silent picking is the failure mode; transparent picking is the discipline.

Common Prompt Mistakes

Output ratio is an enum, not a free-form value — and anamorphic is a style register, not an output dimension. Output aspect ratio is a hard, enumerated platform spec — Kling 3.0 emits 16:9 / 9:16 / 1:1 and nothing else. "Anamorphic" is a cinematography register (anamorphic lens flares, letterboxed compositional read, >2:1 framing aesthetic) that the model can render within a 16:9 output. "16:9 anamorphic" written as a single phrase in the prompt body is incoherent — pick one. Output ratio belongs in the header (and must be one of the enum values for the chosen model — check higgsfield model get <model> or the MCP models_explore equivalent before assuming). Anamorphic style cues belong in the Look line ("anamorphic-style flares, letterboxed composition") as a style request, not as an output dimension.

Negative constraints: For a comprehensive list of artifacts to avoid (floating limbs, face warping, flickering textures, etc.) and the prompt phrasing to prevent them, see ../shared/negative-constraints.md. Always check the relevant categories for your prompt type.

Before You Iterate — Is the Miss Systematic or Stochastic?

At a ~1.5% video / ~1% image acceptance bar, most misses are variance, not a broken prompt. Serial single-variable iteration is the right tool for a systematic miss — the prompt is genuinely wrong. Run it on a stochastic miss and you're "fixing" a prompt that was already right, burning credits to re-roll the same dice one at a time. So decide the fork before you touch the prompt:

• Are the misses all failing the same way? (identity drifts every time, wardrobe contaminates every time, the cut count is always wrong) → systematic. The prompt is wrong. Iterate it, one variable at a time (next section).
• Are they failing in varied ways, with the occasional near-hit? (performance flat on one roll, camera off on another, physics odd on a third) → stochastic. The prompt is right; the roll wasn't. Stop touching the prompt. Lock it, fire a batch, and cull. (When only two takes exist and both failed, the troubleshoot Retry Ladder reads a second failure as over-packing instead — OPEN, no default: ../higgsfield-troubleshoot/SKILL.md § Stop-Rule Ladder, ../shared/house-rulings.md P1-2.)

You don't have to eyeball this. The ledger already classifies every reject as structural or stochastic, and ratio <project> prints a verdict per shot tag: iterate (structural-dominant), batch+sel (stochastic-dominant), mixed, or low-n. Below five logged rows (LOW_N_THRESHOLD) the split is noise — the ledger stays silent and you call it by eye. Read the verdict at the decision point; don't iterate against a batch+sel tag.

Batch-and-Select (Variance-Harvesting) — Not the Same as Stylistic Fan-Out

When the verdict is stochastic, the move is variance-harvesting: hold the same locked prompt constant, roll N at once (grid generation / Batch Size in Cinema Studio; DoP Lite for cheap rolls — not in the API catalog, 2026-09-26 — verify in the live UI), and cull to the keeper. This is the opposite of the stylistic-fan-out exception in the next section — that varies N different looks; this rolls N identical attempts because the prompt is right and only the dice are the problem. They read alike and are economically distinct: fan-out explores, harvest exploits.

The cull rubric — how to pick the keeper from a batch. Batching is worthless without a disciplined select. Don't pick "the prettiest"; select against the falsifiable success criteria you locked before generating:

1. Hard-gate on the invariants first. Identity, wardrobe, cut count, text legibility, named physics anchors — any roll that fails one is out, however nice it looks. (These are your structural failure modes; a batch can't fix a structural miss, only dodge a stochastic one.)
2. Among survivors, score the stochastic axis that was failing — the performance, camera, or composition you were re-rolling for. Best one wins.
3. One keeper, log the rest. Mark the winner kept; log the culled rolls rejected with their real reject_reason so the denominator stays honest and the verdict keeps sharpening. A harvested batch is correctly logged as one prompt_hash, N rows, one keep + N−1 stochastic rejects.
4. If the whole batch fails the hard gate, the miss was systematic after all — stop harvesting and go iterate the prompt.

The Iteration Rule — Change One Variable at a Time

When a prompt is close-but-not-right and you're about to regenerate, change exactly one variable per attempt. Subject detail, composition, motion behavior, lighting, or style — pick the one that's wrong, change only that, regenerate.

Why it matters: if you change two variables and the result improves, you don't know which change drove the improvement. If the result regresses, you don't know which change broke it. Either way you've spent a generation and learned nothing about the prompt. Single-variable iteration gives every regeneration a clean cause-and-effect signal — you keep what works, drop what doesn't, and converge on the right prompt fast.

The exception: once the prompt is locked and you're varying purely for stylistic exploration (e.g., five lighting variants of an already-approved scene), batching changes is fine. The rule applies during refinement, not during fan-out. (Don't confuse this stylistic fan-out — N different looks — with variance-harvesting above, which rolls N identical locked prompts to beat a stochastic miss. Both batch; only one changes the prompt.)

Workflow:

1. Generate the baseline.
2. Identify what's wrong — pick one specific thing.
3. Change only that variable in the prompt; leave everything else untouched.
4. Regenerate.
5. Compare against the baseline — did the targeted change move the result the way you expected?
6. Lock that change. Identify the next problem. Repeat.

If you find yourself wanting to "fix everything at once," stop and ask which fix matters most. That one goes in this regeneration; the rest wait their turn.

Prompt-window hygiene

Iteration also accumulates clutter — old prompt edits that no longer apply, stale reference images attached from earlier shots, contradictory clauses layered atop one another, prompts that have grown so long the model loses the load-bearing pieces inside the noise. Four hygiene patterns from production practice:

• Delete obsolete prompt blocks. When a previous prompt section was about the sticky-note prop but you've moved to character generation, delete the sticky-note block. Otherwise the model occasionally pulls the stale prop into the new generation.
• Editor-adds-atop-existing-prompt creates contradictions. When edits accumulate by appending rather than replacing, the prompt collects contradictory clauses (tight close-up from the old version, wide establishing from the new one). Symptom: output degrades into model-confusion artifacts. Counter: when iterating, replace the relevant clause in place rather than appending a new one.
• Remove stale reference images. When assets change between shots (the Polaroid was on the fridge in shot 1 but pulled off in shot 2), remove the now-stale reference from the prompt window so the model is not still trying to place it.
• Prompt-overload sanitize pass. When the prompt has grown beyond ~2-3k characters from accumulated detail, ask Claude (or whatever prompt- construction surface you use) to optimize / study the context / sanitize the prompt — consolidate redundant clauses, drop now-obsolete qualifiers, preserve the load-bearing structure. The sanitize pass keeps the prompt inside the cap and inside the model's effective attention window.