Higgsfield Seedance

Higgsfield Seedance Director

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.

• The filter is an LLM reading full-scene intent, not a keyword blacklist — describe a SCENE, not a subject; fix the voice first →
• Instant fail (<10s) = filter rejection; delayed fail (>30s) = infra/complexity — never regenerate an instant fail unchanged →
• Six slots, in order: Camera + Subject + Action + Setting + Style + Lighting; missing 3+ slots is where flags come from →
• Empirical prompt-craft laws: 50–80-word attention sweet spot (front-load the load-bearing element), name a director/lens not "cinematic", "fast" degrades motion, no negative prompts in the body, unidirectional motion chains + named camera endpoint + detail scale follows shot size →
• Five prompt modes: Reference-Based / Continuation / Expand Shot / Edit Shot / Transformation — pick the mode before writing →
• [OFFICIAL] block scaffold for production prompts: SCENE CONTEXT → … → POSITIVE LOCKS, distributed style on standalone briefs (connected shotlists glue the compiled Style Prefix verbatim instead), FOV in degrees only, CAMERA block 3rd, cut ladder oner / CUT n / timed / freestyle →
• [FIELD] 13-project corpus calibration: word length scales with register (218w → 2,059w medians — the 50–80w sweet spot is single-shot-only), video briefs hand-authored (enhance_prompt off), Style Prefix = per-project constant compiled into home blocks →
• [FIELD] Three "helpful-instinct" drift sources, each with a standing lock: environment invention (#1, above character drift), character-height equalization, scale drift on wides →
• Build-safe construction for crowds/destruction/creatures: evacuated cities, contained fights ("stays at the sea surface"), the safe benchmark scene →
• Extend an existing clip: attach it as a video reference + open with "The scene continues." — match source resolution AND duration; chain cap ~2 (hard 3), then re-anchor from ORIGINAL references →
• This file is Seedance 2.0. For 2.5 — four modes incl. video_edit / video_extension, 4–30s, 30/10/10 references, first-last frames in omni_reference, up to 1080p (no 4K, no genre) — use ../higgsfield-seedance-2-5/SKILL.md
• Tutorial-demonstrated patterns (reference-role vocabulary incl. VARIETY reference, SCREEN REALISM + duration-match composites, 60:30:10 grade, red-arrow prop annotation): PRODUCTION-PATTERNS.md in this directory
• [OFFICIAL] Feature-film pipeline (asset construction, per-scene GEO SPATIAL LAYOUT, the position-fixing first second, dialogue construction, ban dictionary, the 10–15 iteration rule, crowds / giants / thresholds): HELL-GRIND.md in this directory
• Performance — objective, obstacle, tactics, beats, subtext, eye life, the acting master profile: ../higgsfield-acting/SKILL.md
• Hard engine rules (age-blind, exit-frame = cut, off-screen = nonexistent, no reflections, ≤3 tracked characters, double-contrast cuts) + high-risk shot table: ENGINE-RULES.md in this directory
• Render failures (the take came back wrong, not flagged) — 15 named failures with symptom · mechanism · counter, incl. action-reversal fill, filler-babble, truncated action, mimed manipulation, orphan limbs / the solo third hand, walking + cross-shot camera speed, fights as separate clips — plus a self-repair checklist: FAILURE-MODES.md in this directory
• [FIELD] Bake it into the asset: a property that drifts however well it is written (a lens character, a grain, a colour cast) moves one step earlier, into the plate — then, in one studio's practice, leaves the video prompt entirely (OPEN: Hell Grind keeps the look in the prompt too, P2-6); bake only what should stay constant →
• [FIELD] Depth map as a reference role — greyscale, light = near; a geometry input that stops a space re-planning itself between shots, carries no style →
• Contested doctrine (character-text volume with a reference attached, anamorphic words vs a baked plate, NO BGM vs No music., the stop numbers) is recorded — not silently resolved — in ../shared/house-rulings.md; label meanings in ../shared/provenance.md
• Reference roles: Character / Last-Frame / Environment / Prop — role determines what the prompt may re-describe →
• Working modes: Exploration / Continuation / Bridging / Repair (distinct from prompt modes) →
• Layer 1 briefing vs Layer 2 production prompt — never paste Layer 1 into the prompt box →
• Native 4K is available in mode=std only; mode=fast (Seedance 2.0 Fast) caps at 480p/720p — in Cinema Studio the model is still capped at 1080p →
• Always preflight: python3 scripts/seedance_lint.py --preflight --model seedance_2_0 "<prompt>" — enums come from ../../specs/model-specs.json (fast+1080p/4K and Kling 21:9 are auto-caught) →
• 480p drafts validate the prompt, NOT the take — no seed param; pin Hero Frame + start/end frames to carry a look →
• ZH prompts: hard 1,800-char cap; ZH antislop list enforced by the linter →
• Flagged prompt → rewrite playbook per linter rule, then voice pass →
• Repeated flags → full loop-breaker procedure + LOG THE OUTCOME (--confirmed / add-quality) →

Use this skill whenever the user wants a Seedance 2.0 / Seedance Pro prompt, OR whenever a Seedance generation has been blocked, flagged, or silently failed. This skill's job is to stop credit waste on filter rejections.

Engine rules (read with this file): the hard rendering constraints of the Seedance 2.0 engine — age-blind characters, exit-frame = implicit cut, off-screen = nonexistent, no reflection shots, ≤3 tracked characters, the double-contrast cut rule — live in ENGINE-RULES.md in this directory, together with the high-risk shot table (reflections, same-character doubles, crowds, text rendering) and its mitigations. This SKILL.md is the EN-director profile of that rule core; the ZH-house and bilingual-JSON profiles (../../docs/Seedance 2 Skill.md) obey the same core. Flag high-risk shot types at authoring time — never silently break a rule the project's hero image happens to conflict with.

Production patterns (sibling reference): patterns demonstrated working in Higgsfield's own Seedance-4K film tutorial — reference-role vocabulary, coordinate blocking, video-reference screen composites, prompted imperfection, 60:30:10 grade — live in PRODUCTION-PATTERNS.md in this directory, labeled [DEMO].

The Filter Model — Read This First

Seedance 2.0's content filter is not a keyword blacklist. It is a language model that reads the full prompt as a single scene and judges intent and context. Most users burn hours swapping individual words — that loop does not work.

The filter compares two things:

• A prompt that reads like a filmmaker describing a shot → tends to pass.
• A prompt that reads like a note to a friend → tends to fail.

A word that looks sensitive in isolation can sit inside a well-constructed cinematic prompt without issue — the filter reads the full picture. A prompt with no picture to read (no setting, no visual purpose, no narrative logic) gives the filter nothing to work with, and it errs on the side of caution.

Practical rule: the prompt must describe a scene, not a subject. Fix the voice first, then fix the words.

Instant Fail vs. Delayed Fail — the Diagnostic

This single heuristic saves time on every failure:

If the user is seeing instant fails in a loop, it is a filter issue — never a GPU issue. Stop them from regenerating before the rewrite.

The Seedance Prompt Formula

Every Seedance prompt should hit these six slots, in this order:

[Camera movement] + [Subject] + [Action] + [Setting] + [Style] + [Lighting]

All six are technically optional — but a prompt that includes all six almost never gets flagged, because the filter has full context to interpret every word. A prompt missing 3+ slots is where flags come from.

Minimum viable Seedance prompt

Slow dolly-in on a figure in a dark overcoat standing alone at the end of a rain-slick alley. Cold teal shadows, single practical streetlamp, shallow depth of field.

Camera ✓ Subject ✓ Action ✓ Setting ✓ Style ✓ Lighting ✓ — all six slots, ~30 words, passes the filter because the scene is fully legible.

Prompt-Craft Laws

A set of Seedance-2.0-specific prompt rules. These are empirical — practitioner A/B findings that are plausible given the architecture but are not in the official model spec. Treat them as strong heuristics and let the repo's iteration discipline (../higgsfield-prompt/SKILL.md § The Iteration Rule) confirm them on your own material, rather than as guaranteed model behavior.

Length and order — the attention model

Seedance reads the prompt left-to-right with diminishing attention weight. The first sentence carries the most influence; by the third sentence you are in "detail territory," where the model stops treating elements as primary instructions and starts sampling them diffusely.

• Sweet spot: 50–80 words (short-form regime). A 70-word prompt reliably outperforms a structurally identical 200-word version of the same scene — more words past ~3 sentences buys diffusion, not control. (Block-scaffold production prompts are the other regime: § Official Prompt Architecture.)
• Structure in three sentences: ① subject + action, ② camera + style, ③ constraints / positive locks.
• Lead with the single most load-bearing element. When a shot lives on its subject, the subject opens the prompt; when it lives on a camera move, the move opens it.

Relationship to the two length numbers. This 50–80-word figure is the coherence optimum. The >180-word figure in § Pre-flight Linter is a different axis — the filter/encoder risk ceiling (>220 often hard-fails the text encoder). 50–80 is where to sit; ~180 is where it starts to break. They don't conflict.

Relationship to the six-slot formula. The six slots guarantee the filter sees a complete scene (presence). The attention model governs weight (order

• length). Keep all six slots present, but the slot list's camera-first ordering is a completeness checklist, not a mandate to open with the camera word when the shot's identity is the subject.

Name the thing — kill empty adjectives

cinematic, epic, beautiful, high quality, amazing are high-frequency labels attached to an enormous range of training footage — dark thrillers, bright rom-coms, nature docs all read as "cinematic" — so the model samples a broad, diffuse distribution and they move the output toward nothing in particular. Don't just delete the slop word (Voice Rewrite §6) — substitute a named, narrowly-trained referent:

Positive form of ../higgsfield-prompt/SKILL.md § Anti-Slop Vocabulary. The lens-spec row is short-form, single-shot doctrine: for a sequence whose location plates are generated with the lens already in them, one studio drops the optics words from the video prompt (§ Bake it into the asset) while Hell Grind keeps the look in both the plates and the prompt — OPEN (../shared/house-rulings.md P2-6).

Official override on director names. Higgsfield's own prompt-writing skill forbids director names, signature-work references, and equipment model names outright (see § Official Prompt Architecture — the Block Scaffold → Measurable-language rules). The director-substitute trick above is an empirical short-form fallback; in block-scaffold prompts, describe the look in observable terms instead — "centered symmetrical framing, pastel palette", not "Wes Anderson symmetry".

"fast" is the highest-degradation keyword

Combined with complex action or camera movement, fast is the single worst-degrading keyword. The temporal branch already runs multiple high-velocity calculations when motion is layered; fast asks all of them to run at maximum velocity at once. Two competing fast elements jitter; three compound into error that's hard to salvage.

Fix: describe the physics, not the speed. feet striking hard, each stride at full extension, arms pumping at 90 degrees produces the perception of speed with no degradation. One element can carry speed — just not all of them simultaneously. (Same family as Voice Rewrite §3 — describe physics, not emotion.)

No negative prompts in the prompt body

Seedance has no negative-embedding architecture for the prompt text — every token is read as a positive instruction. negative: jitter, bent limbs gets parsed as scene description the model tries to render (noise), not as a constraint, and makes the output worse.

Use positive constraint statements — direct declarations of what must be true:

Face stable. Limbs anatomically natural. Consistent lighting, no flicker.
Body proportions consistent throughout.

Scope: this is about the Seedance prompt body, and the target is negative: list syntax / bare negation lists — not every "no" token. A short lock tail inside a positive declaration ("Consistent lighting, no flicker"; the Style Prefix's "Photorealistic — no 3D render") is fine and field-proven across the harvest corpus. It does not override the Higgsfield UI's dedicated negative-prompt field (which some image models expose and ../../vocab.md § Composition Vocabulary uses). The same positive-only requirement is already documented for Cinema Studio 3.0 in ../shared/negative-constraints.md, which also carries the test for when a short, specific ban is correct (§ Where a ban is still correct — is the model's default already the failure?).

Ambiguous verbs — the homograph trap (v1.10, Peter's find 2026-07-14)

If a word has a plausible second reading, Seedance may take it. The observed case: "wind tearing at her coat" — meant as fabric pulled violently; the model sometimes reads tearing as ripping (fabric shredding) or tearing up (crying), and the shot changes accordingly. This is not covered by any known prompt guide — treat it as a first-class law:

Before a verb ships, ask: is there a second physical thing this word can look like? If yes, replace it with the phrasing only ONE thing can look like.

• wind tearing at her coat → wind whipping violently at her coat / her coat flutters violently in the wind
• Seed homograph list (grow it whenever a generation misreads a word): tearing (rip / cry) · shoot (fire / film) · duck (crouch / bird) · bolt (run / lightning / hardware) · draw (pull / sketch / weapon) · wave (hand / ocean) · charge (run at / electricity) · rock (sway / stone) · drop (fall / droplet) · fire (flame / shoot / dismiss) · strike (hit / match / lightning) · break (shatter / pause / dawn) · pound (hammer / heartbeat) · snap (break / photo / fingers).
• The list is a seed, not the rule — the rule is the self-check, which generalizes to any word forever.

Community v3 cherry-picks (Joey drop, audited 2026-07-14)

Adopted (genuinely absent from this skill until now):

• Camera on the shadow side, with a stated operator axis. Place the camera on the shadow side of the key light and say where the operator stands/moves — light wraps toward the lens and faces keep dimension.
• Detail-on-wide ("snake cam"). 84° low-angle placed hard against a small foreground object — detail-shot intimacy without losing the wide's context.
• Intimate wide. 63–84° on a close face instead of a long lens — presence without compression; the room stays in the frame.
• Prompt-reset heuristic. When iterations are getting worse, stop stacking fixes: strip the prompt back to subject + action + camera and re-add only what's necessary. Density is a bell curve; past the peak you can't tell which element the model dropped.
• Canonical-over-plate. Every subject keeps its own identity reference even when it is visible in the environment plate — the plate carries the world, the canonical ref carries identity; never let a plate double as an identity source.
• Contrast curve stated three ways. When the grade matters, state it as tonal curve + specular removal + named grade — one phrasing alone drifts.

Rejected (was: flagged, test day pending): their worldbuilder puts the camera block at the BOTTOM ("at the top FOV fights identity data") — this contradicts both this skill's CAMERA-3rd-position rule and their own seedance skill. Resolved 2026-07-26 by field evidence instead of a test day [FIELD — 13-project community harvest]: across ~4,000 harvested production prompts from 9 creators, the CAMERA block sits mid-document in every final prompt — never at the bottom. CAMERA-3rd stands; the bottom-position claim is dropped.

Motion-prompt laws (dramaclaw production corpus, audited 2026-08-09)

[EMPIRICAL — dramaclaw production corpus, Seedance] — practitioner findings earned in dramaclaw's Seedance production work. The craft is model-agnostic i2v motion-writing rather than a Seedance spec; same epistemic status as the rest of this section (strong heuristics — confirm on your own material).

• Unidirectional motion only. A short action that finishes early leaves the model with seconds of clip to fill, and it fills them by reversing the action — the character walks forward then steps back, leans in then pulls away. Chain 2–3 connected actions in the same direction so the motion spends the whole clip; a deliberate there-and-back is two shots, never one prompt. (Failure face: FAILURE-MODES.md § Action-reversal fill.) Scope (../shared/house-rulings.md P2-1): this chain is for low-complexity, same-vector motion whose job is to fill the clip — walks to the window, pulls the curtain, leans into the glass. A multi-step process that reverses direction on the way to a peak (reaches in, pulls out, winds up) is the case ../higgsfield-acting/SKILL.md § States, not transitions opens already in the state instead; an object that must visibly change under the hands takes the causal chain in FAILURE-MODES.md § Mimed manipulation. Pick by what the shot is for.
• Name the camera endpoint. A camera move needs a destination, not just a name — say what the frame shows when the move finishes ("slow dolly-in, ending on her hands wrapped around the cup"), not only the move's name. A move that runs out of instruction before it runs out of clip drifts or reverses — the camera face of the unidirectional law.
• Detail scale follows shot size. Close-ups earn micro-detail (fingers tightening, a jaw flex); wides earn broad arcs (crossing the courtyard, the crowd parting). Cross-matching — micro-detail written into a wide, or a broad traversal written into a close-up — is unrenderable at that shot size and degrades the whole clip. (Detail inside a wide is a composition problem, not a prompt-detail problem — see the snake-cam cherry-pick above.)

Already-covered siblings (cross-links, not new rules)

• Compound camera move (dolly in while panning left) → jitter at the transition because the model executes the two vectors in sequence. Use one primary move + one texture modifier (slow dolly in, slightly handheld). Full treatment: FAILURE-MODES.md § Multi-motion camera overload.
• Image-to-video subject drift → re-describing what's already in the source image gives the model two competing inputs for one subject; reconciliation introduces drift. Keep an I2V prompt to motion + camera only. See ../higgsfield-prompt/SKILL.md § Image-to-Video (I2V) and ../higgsfield-seedance-vfx/references/first-frame.md (the start_image handoff). This rule is for an image that is frame one; an image attached as an identity reference is § Reference-Based, where how much matching identity text may ride beside it is OPEN (../shared/house-rulings.md P1-1).

Official Prompt Architecture — the Block Scaffold

[OFFICIAL — Higgsfield prompt-writter.skill, 2026-07] — Higgsfield ships its own Seedance 2.0 prompt-writing skill with the Seedance-4K release. This section is that doctrine, reconciled with the rest of this file. Where the two disagree, the official rule wins inside block-scaffold prompts; the empirical rules elsewhere in this skill remain the short-form regime.

Two regimes, not a contradiction. The six-slot formula and the 50–80-word sweet spot (§ Prompt-Craft Laws) govern short-form single shots. The block scaffold is the production regime — multi-shot, reference-heavy, high-control work — where structure replaces the word cap: write densely where control matters, sparsely where it does not, and say each important thing once. The pre-flight linter detects this regime automatically (canonical block labels / shot markers) and suspends the short-form word caps — force it with --regime block if detection misses — while every structural lint rule (shot counts, beat sums, handle declarations, enum checks) still applies in full. Likewise, the Voice Rewrite instruction to put a Style & Mood clause up front is the short-form filter pass — in a block prompt the filter gets its full scene from SCENE CONTEXT, LOCATION MAP, and LIGHTING instead, and style is distributed (below).

Block order

Write blocks in this order, using only the blocks the shot needs:

SCENE CONTEXT
ACTIVE REFERENCES
LOCATION MAP
FIRST FRAME / BLOCKING
FORMAT MODE
OPTICS
CAMERA
ACTION
PERFORMANCE      (when acting matters)
PHYSICS
LIGHTING
COLOR GRADE      (when the grade is strong / stylized)
WARDROBE         (when costume matters)
AUDIO
STYLE            (technical-style suffix)
OUTPUT SETTINGS  (when format must be pinned)
POSITIVE LOCKS

Logic: context and references first, then space and timing, then action and physics, then descriptive style in its home positions, then a technical suffix, locks last. A naturalistic single take may drop COLOR GRADE, WARDROBE, and OUTPUT SETTINGS entirely and fold those notes into LOCATION MAP / LIGHTING.

Distributed style — the standalone-block rule

In a standalone block-scaffold prompt there is no style-prefix block at the top; the prompt always opens on SCENE CONTEXT. Each style aspect lives in the block that governs it: light → LIGHTING; color → COLOR GRADE, or folded into LOCATION MAP + LIGHTING for a naturalistic look; lens / optical character → OPTICS; skin realism and acting → PERFORMANCE; format / grain / fps → the STYLE + OUTPUT SETTINGS suffix just before POSITIVE LOCKS. Descriptive style sits in the body next to what it describes; technical style sits as the end suffix; nothing style-related opens the prompt.

Connected-shotlist carve-out [FIELD — 13-project harvest]: in a multi-scene shotlist project, the field-proven shape is the opposite — a per-project compiled Style Prefix glued verbatim to the top of every scene prompt (edit once → changes everywhere), then the scene body. Consistency across 25 separately-generated scenes outweighs distributed elegance there. That regime is owned by ../higgsfield-shotlist-director/SKILL.md § Per-scene prompt law and ../../templates/seedance/global-style-prefix.md; this section governs standalone prompts only.

FOV anchors + the CAMERA-3rd-position rule

In prompt text, state field of view in degrees from these discrete anchors only — never millimeters, never in-between values (not "23°" — use 18° or 29°):

In a multishot, set FOV per segment and add "no drift mid-segment".

Place the CAMERA block in the 3rd position of the prompt's core layers (subject → action → camera → style → constraints). Moved to the end, FOV gets ignored; moved to the front, it conflicts with identity.

Measurable-language rules

• Positive phrasing only — official confirmation of the empirical no-negative-prompts law (§ Prompt-Craft Laws).
• Speeds in km/h — "moves at 40 km/h", not "fast" (which is also the highest-degradation keyword, § Prompt-Craft Laws).
• Atmosphere in % / meters — "fog density 40%", "haze visible at 15 meters"; build it in steps across shots (20% → 40% → 60%).
• Giant scale via human-height comparison — "as tall as four humans stacked head to toe", not "huge" or "three meters".
• Left/right is always from the camera.
• Emotion through muscle movement, not labels — same rule as Voice Rewrite §3; the muscle-level extreme is ../higgsfield-facs/SKILL.md.
• White balance in Kelvin, fixed within a scene: 3200K / 4000K / 5600K / 8500K.
• Masses and sizes in real units for physics [FIELD] — the PHYSICS block writes weights in so gravity reads correctly: "the spider is 10–12 centimeters, 50–70 grams — it falls gently; the man is 70–80 kg — the drop lands hard." Material behavior gets the same treatment ("the liana stretches slightly under load and creaks like fibrous wood").
• Causal prop interaction [FIELD] — objects never move on their own; every prop event needs a visible physical cause with event order: "the cup tips only from visible sleeve contact" · "a button press is a full mechanical event — contact, 2–3 mm of travel, click, spring-back; the screen lights only AFTER the click."
• No director names, signature works, or equipment model names — they get ignored or break complex moves; describe the look instead. (Overrides the empirical director-substitute in § Prompt-Craft Laws for this regime.)
• English prompts only (for the historical ZH exception, see § Multi-Language Prompt Workarounds).

POSITIVE LOCKS

A lock is a short hard fixer placed next to what it protects — "headlights stay glowing in every shot". The POSITIVE LOCKS block closes the prompt: continuity (characters, props, environment identical across cuts) plus a single positive restatement of critical info. This is where the positive constraint statements from § Prompt-Craft Laws live in a block prompt.

Three named "helpful-instinct" drift sources [FIELD — 13-project harvest] — the model's own instincts, each needing a standing lock:

1. Environment invention — the #1 drift source, above character drift. Whenever a location reference is attached, the model "helpfully" widens rooms and adds furniture it wasn't shown; multiple creators called this the single largest source of drift between clips. Standing lock: "the set contains only what the reference shows — no added furniture, rooms, or geography beyond the reference", plus explicit absences where they matter ("nothing on the floor, nothing on the bed").
2. Character-height equalization. With two people in frame the model drifts toward equal heights. Write real heights into every prompt with 2+ characters ("she is 165 cm, he is 178 cm") so relative scale never floats.
3. Scale drift on wide shots. A human anchor shrinks to a speck across cuts. Lock it: "the rider stays the calm human-sized anchor — never shrunk to a tiny distant dot." (Adapted: the harvested lock named the anchor by an age word; rewritten age-blind per ENGINE-RULES.md rule 1.)

Cut-format ladder

Four precision levels — points on a scale, pick the one the shot needs:

1. Oner — "one continuous shot, the camera does not cut on its own."
2. Sequential cuts, no timecodes — CUT 1 … CUT 2 … CUT 3, described in order, when cuts matter but exact timing doesn't.
3. Timed multishot — explicit cuts at stated seconds (1.0s HARD CUT), when beats must land on a clock.
4. Freestyle b-roll — don't lock cuts; let the model find angles.

Whenever cuts are specified (timed or not), add: "cuts only at the specified points, the camera does not cut on its own." Cut vocabulary: HARD CUT, SMASH CUT, MATCH CUT, INSERT CUT, REVERSE CUT, WHIP CUT; fades/crossfades only if explicitly requested. This ladder is the resolution of the pick-a-side anti-pattern in § Output Format (per-second labels inside an intended oner read as cut instructions), and timed beats must still sum to the declared duration (4–15s) per the runtime arithmetic there.

Tag naming + minimal reference text

• User-specified tags verbatim; otherwise load-order @image1 @video1 @audio1 — consistent with § Reference Roles → Per-Image Role Convention.
• The @TAG: reference line = role/build + current state + unique visible features + action-critical details + voice (only if it has a line) + "100% matches the reference". No age. The official skill's own template opened this line with an age token; engine rule 1 (ENGINE-RULES.md) forbids it in either language, and Higgsfield's feature-film brief gives the reason — the content filter tightens sharply the moment it reads a minor (HELL-GRIND.md § Wording rules). Carry the same information through build, wear, and posture.
• Keep reference character text minimal — long appearance text fights the image and degrades it (same mechanism as the I2V subject-drift rule in § Prompt-Craft Laws). Contested (../shared/house-rulings.md P1-1): Higgsfield's feature-film pipeline pastes the full descriptor, word for word, next to the same kind of reference (HELL-GRIND.md § The core problem). Both are Higgsfield sources and nothing here measures which holds identity better. Settled on both sides: on a start frame (I2V) — motion and camera only; with no reference — the descriptor is the only identity carrier and goes in verbatim every time; whatever identity text you write is fixed wording that never contradicts the reference.
• State critical details in words anyway — small text, logos, colors — even when visible in the reference; the model can drop them.
• Never place an @tag in a shot where that object is not present — the model will force it into frame.

Naming the tags themselves [FIELD — Higgsfield Studio, ONEIRIC breakdown, 2026-08-13] The bullets above say what a tag line contains; on a project of any size the name needs a convention too, and the rule behind it is one element, one name. Without it a project grows duplicates — the same couch living under three names — and nobody can tell which reference is the real one:

@loc_ON_dorm_commonroom_front_s2     type + project + name + [angle] + scene
@char_ON_Rudy_s2_v1                  type + project + name + scene + version
@prop_ON_pizza                       type + project + name

• The scene suffix ties an asset to where it lives, so two dressings of one room, or two versions of a location across a time jump, cannot collide.
• The version suffix appears when an asset's pixels change — a character's state change, or a plate or diagram regenerated — and a changed asset is a new asset with a new name, never an overwrite. One character in a dorm room and the same character in a hospital bed are two assets of one man. The identity discipline behind that split lives in ../higgsfield-soul/SKILL.md § The Untouched Base; this bullet is only the naming half of it. A location carries a version too once its plate is regenerated — the staging-side form (@loc_…_v2 beside @staging_…_v2) bumps only the tag whose image changed.
• Tags are arbitrary strings, so mixed case is safe and consistency is the only rule. Pair related names visibly (a location and its staging reference sharing a stem) so it reads at a glance which assets belong together — see ../../templates/seedance/staging-reference.md § Tag naming for the staging-side form.

Context isolation

Every generation is a blank slate with no memory of previous shots. Never carry in scene numbers, script headings, prior-scene summaries, unused tags or characters, or "as above / continues" phrasing. This is why the Continuation Prompt Formula (below) demands a verbatim identity re-paste rather than a reference back to the earlier prompt.

Special protocols

• 4-mechanism extreme-FOV multishot stack (8° / 107°): ① sequence-wide identity lock (single location reference across all beats), ② LENS LOCK opener — explicit FOV phrase starting each beat, ③ LENS CHECK closer confirming FOV at the end of each beat, ④ color via material + light, not a list. All four or extreme-FOV multishots break down after 2–3 beats.
• Whip-pan needs ≥0.8s of blur travel — under 0.8s it renders as a hard cut without blur (settled → 0.8s WHIP → settled).
• Mixed real-time / slow-mo: hard cuts only between speed modes; each shot is one speed start to finish.
• Anti-impact locks for cracks/breaks: "crowd PRESSES, not strikes" · "fracture originates from edge stress, not center impact" · "no impact point — pressure-based crack" · sequential timing edge-to-center, not radial from a point.
• Observation pattern (hidden-camera effect), all three at once: foreground occlusion over 20–30% of frame + atmospheric haze between camera and subject + distance vantage at 8–12°. Change the occlusion type between beats; keep the vantage single.

For these protocols applied on real footage — per-segment LENS LOCKs, timed SMASH/MATCH cuts, screen composites — see PRODUCTION-PATTERNS.md in this directory.

Field calibration — the 13-project production corpus

[FIELD — community harvest 2026-07-18: 13 shared Higgsfield projects, 9 creators, ~4,000 production prompts pulled via API with full params]. What the corpus confirms and calibrates about the block scaffold:

• The scaffold holds platform-wide. Every project — photoreal adventure, K-drama romance, broadcast-TV drama, anime, stop-motion folklore — runs the same Style-Prefix-plus-Constraints-plus-variable-SHOT structure with the same block anatomy. "SHOT" markers appear in 95% of the flagship project's 1,240 Seedance prompts; 15s is the dominant duration (they generate long multi-shot clips and cut the best seconds).
• Word-length ladder by register. Median Seedance prompt length tracks register and ambition, not a fixed cap: tech-demo 218w → broadcast-TV drama 538w → commercial 779w → genre anthology 955w → adventure film 1,433w (p90 2,648) → stop-motion emotional drama 2,059w. The 50–80-word sweet spot (§ Prompt-Craft Laws) is single-shot doctrine; production multishot briefs live an order of magnitude above it. Length scales with performance complexity — never truncate a reaction arc to be neat.
• Register contracts the template. Photoreal keeps every block (PHYSICS with real masses, pore-level PERFORMANCE); stylized work (anime, stop-motion) drops PHYSICS/skin realism and keeps SHOT beats + continuity + a hard medium lock. See ../higgsfield-style/SKILL.md § Register Poles.
• The Style Prefix is a per-project compiled constant. In connected shotlist projects the corpus ships it verbatim at the top of every scene prompt (the Style-Prefix-plus-Constraints-plus-SHOT structure in the first bullet — that is the delivered form, not just an authoring note). Standalone block prompts instead distribute those aspects into their home blocks (§ Distributed style — the standalone-block rule). Which form ships is decided by the workflow, not by taste — see ../higgsfield-shotlist-director/SKILL.md § Per-scene prompt law.
• Video prompts are hand-authored. enhance_prompt was absent/off on every harvested Seedance job but ON for 1,022 image jobs — let the enhancer expand image prompts, never the video brief.
• Platform-layer params observed on Seedance jobs (Higgsfield surface, not necessarily the raw model API): multi_shot_mode: "custom" (the timed multishot mode), genre: "auto", speedramp: "auto", mode: "std", bitrate_mode: "high", generate_audio: true, 21:9 at 4K.
• Iteration economics (TESTS-first culture, 65–100 generations per kept shot, five-bucket folder discipline): ../../production-benchmarks.md § Community-corpus anchors.

Seedance 2.0 Prompt Modes

Seedance 2.0 exposes five generation modes that each take the six-slot formula but apply it to a different starting point. Picking the right mode is upstream of prompt writing — the same sentence will produce different results in different modes, because each mode reads the prompt as a different kind of instruction.

Reference-Based

The prompt builds a scene around a source image that carries the visual identity — character, wardrobe, palette, sometimes composition. The prompt's job is NOT to re-describe what the image already shows; it's to place the subject into a new action, setting, or motion context. This is the workhorse mode for any sequence that needs a consistent character across varied shots. (How much identity text may still ride next to the image is contested — ../shared/house-rulings.md P1-1.)

[Source image role: "as the main character" / "as the starting frame"].
[Action the subject performs]. [Environment and atmosphere if not visible in source].
[Camera movement]. [Lighting cue if different from source].

Continuation

The prompt extends a prior Seedance generation forward in time, picking up at the final frame of the previous clip. Identity, wardrobe, environment, and emotional state all carry over. The prompt should describe what happens NEXT — never what just happened. For the full five-rule construction pattern, see the Continuation Prompt Formula section directly below.

[Continuing from prior clip]. [New action that follows from the last frame].
[Camera direction for the continuation]. [Any state cha