Higgsfield Shotlist Director

Higgsfield Shotlist Director

Turn a brief into one connected shotlist — not a pile of separate prompts. The deliverable is a single editable HTML artifact the user opens in a browser, ticks scenes off as they shoot, and comes back to you to revise. This is the artifact-shaped workflow a fully-AI commercial actually runs on: lock a global look once, declare the cast/props/locations once, then emit named per-scene prompts that all inherit both.

This skill is the connected layer on top of higgsfield-seedance. It does not reinvent the prompt grammar — every per-scene prompt obeys the six-slot formula and the Prompt-Craft Laws in ../higgsfield-seedance/SKILL.md, and every prompt is preflight-linted before delivery. What this skill adds is the document: the global Style Prefix, the @-glossary, the per-scene numbering, and the edit-once / per-scene-override semantics that keep 25 prompts in sync.

QUICK FACTS

• Output = one self-contained HTML file (inline CSS/JS, no deps), not loose prompts →
• Three structural layers, top to bottom: Global Style Prefix → @-asset glossary → named per-scene prompts →
• Per-scene prompt law: Style → Characters → Scene → CUT 1..N; each prompt targets 15s; split long scenes as 3a/3b/3c →
• [OFFICIAL] Density heuristic: group rows when ALL of {same cast, same location, one emotional unit, ≤15s, inside length limits}; split on ANY of {location cut, cast change, setup change, performance arc, insert} — don't fragment grief; complexity budget + auto-enrichment defaults for thin briefs →
• Whole-sequence checks before delivery: tempo budget (cut durations sum exactly to runtime; one 6–8s hero hold) + monotony audit (no 3 consecutive cuts sharing shot size AND camera move) →
• Continuity carries exits too: an Off-screen line (exit side + last state) per just-departed character keeps re-entry direction legal →
• Edit-once-propagates: change the prefix once → it changes in every prompt; per-scene override lets one scene break the global look →
• Differentiators over a bare shotlist generator: preflight linter, reference-role lanes, Elements @-auto-attach, failure-mode awareness, acceptance-rate logging →
• English prompt text only (Seedance expects English), even if the user writes in another language →

What you produce

A single shotlist.html (saved to the user's outputs and presented). It is self-contained — inline CSS, inline JS, zero external dependencies — so the user can open it offline and it just works. Structure:

1. Title bar — project name (infer from the brief; "Untitled" if unclear).
2. Global Style Prefix — collapsible block at top, applied to every prompt.
3. @-asset glossary — the cast/props/locations declared once.
4. Scene list — numbered scenes, each with a checkbox (progress saved in localStorage), a one-line scene description, and one or more copy-ready prompt blocks (Prompt 3a · 15s, Prompt 3b · 15s).
5. A short "how to use" note (checkboxes auto-save; ask Claude to revise).

The Style Prefix appears once in the collapsible block and is prepended verbatim to every prompt's copy-block — so the user copies one prompt into Seedance and it works standalone, no reassembly.

The three layers

1. Global Style Prefix

A single style block glued to every prompt in the document — edit it once and it changes everywhere. It locks the film's global look: format/resolution, lighting doctrine, colour ratio, lens/shutter, skin realism, acting register, physics, composition, continuity, frame rate, and audio convention.

Ship the reusable fill-in-the-blanks block from ../../templates/seedance/global-style-prefix.md. Always check the conversation first — if the user pasted a custom prefix, use that one verbatim. Otherwise use the template default.

2. @-asset glossary

Declare every recurring asset once, with a stable @-name, then register each under Elements in Higgsfield with the same name so pasting a prompt auto-attaches the right images:

@hero — main character          @boss — side character
@headphones — product           @sneakers · @bag · @skydancer — props
@kitchen · @stadium · @street — locations
@s_hero — athletic-look hero     @s_hero_wet — sweaty post-run hero
@music_track (audio_1.wav) — motion locks to this beat
@staging_ACME_street_v1 — front-on position reference, attached LAST

A position map in the glossary is always the front-on staging reference (../../templates/seedance/staging-reference.md: outline figures, position only, attached after the photo references), named by that template's § Tag naming (@staging_[PROJECT]_[scene]_[version]; ACME above is a placeholder project). No slot filename goes beside it: slots are assigned by upload order, so a staging reference filed as image_1 would sit in the character's slot. Elements auto-attach finds the right images, but nothing in this repo documents the order it attaches them in — when the staging reference must land last, check the attach order before firing, or attach it by hand after the photo references. A top-down floor plan is never registered or attached — it is an authoring aid whose output travels into the prompt as a written blocking note (../../templates/seedance/top-down-map.md).

The slot→role discipline (@Image1 = character, @Image2 = costume, @Audio1 = rhythm…) comes from ../higgsfield-seedance/SKILL.md § Reference Roles → Per-Image Role Convention. Multi-state variants get their own locked entry (@s_hero_wet), built on purpose up front — asking the model to "sweat him up" later makes it improvise and the face drifts.

Each glossary entry also carries a fidelity grade — a role says what job the asset does, the grade says how much of it must survive into the pixels: full-preserve / partial-preserve (name the parts) / attribute-transfer (name the target it lifts onto) / loose-guide. One word per line is enough (@headphones — product, full-preserve); the grades and their prompt phrasing live in ../higgsfield-seedance-2-5/SKILL.md § Reference Roles → Fidelity. Without a grade, "use @image4 for the coat" silently means whatever the model felt like keeping that day.

3. Named per-scene prompts

Every scene is numbered (1, 2, 3…) and split into named 15-second prompts (1a, 1b, 2a). One checkbox per scene, even when split across 3a/3b/3c.

Per-scene prompt law

Every prompt follows this exact order, top to bottom. (This verbatim-prefix shape is the connected-shotlist regime [FIELD — 13-project harvest]; a standalone block-scaffold prompt instead distributes style into its home blocks and opens on SCENE CONTEXT — ../higgsfield-seedance/SKILL.md § Distributed style. Which shape ships is decided by the workflow: shotlist → this law; single standalone brief → distributed.)

[STYLE PREFIX — full block, verbatim (or the per-scene override)]

Characters:
[Only the characters in this prompt. @names + locked physical descriptors +
carried state — wet hair from the prior scene, strap on one shoulder, same
wardrobe unless it changed on screen.]

Off-screen (only when someone just left):
[Anyone in the PREVIOUS prompt but absent here: exit side + last visible state —
"Bo — exited frame-left, still carrying the crate." Carry for one prompt, drop
after two consecutive absent prompts.]

Scene:
[1–2 sentences. Where, when, and the geo-spatial blocking — where each character
sits relative to the location and to each other. "Hero at the kitchen island,
back to camera; the moka pot is on the left burner."]

CUT 1 — [framing, lens/FOV, camera move]:
[Beat-accurate action: gesture, eye-line, breath, micro-pause; what the camera
does; what the light does; diegetic SFX if relevant.]

CUT 2 — …

Each prompt targets 15s (Seedance generates a fixed-length clip — design the cuts to fill it, don't pad with dead air). Most 15s prompts hold 1–3 cuts — that is the live-action narrative norm; stylized registers run denser by design (3D-animated 6 shots/15s, product montage 8–10 sections with 0.3s macro cuts — ../higgsfield-style/SKILL.md § Style Recipes), and the flash-establish / insert durations in ../higgsfield-camera/SKILL.md § Shot duration by type are what make dense shapes fit. If a scene runs longer, split it across 3a/3b/3c, each its own 15s block with the full Style Prefix and Characters block, continuity holding across them.

Beat-by-beat choreography, not "he dances." Generic motion verbs mean nothing to Seedance — spell the move out: "two crisp head nods on the beat, shoulders rolling back one at a time, a soft knee-dip, a loose finger-snap, finishing on a quarter-spin." (Full pattern: ../../templates/10-dance-music-performance.md.)

The Off-screen line is what keeps re-entries legal. Carried state covers who is in frame; it says nothing about who just left, and re-entry from the wrong side is a classic continuity break the viewer feels before they can name it. One line per departed character buys the next prompt the correct entering side and hand-state for free. [EMPIRICAL — MiniMax H3 skill corpus; re-derived]

Match-cut via a repeated anchor action. When independently-generated scenes must cut together, end and begin neighboring scenes on the same gesture (the ear-cup tap) — the reused motion lets them "cut on action" most of the time. When those clips then sit on one timeline, plan the unifying finish pass and cut placement per ../higgsfield-audio/SKILL.md § Cutting to music.

Prompt density — grouping shot rows into 15s envelopes

[OFFICIAL — Higgsfield shotlist-builder + seedance-2-pro-director skills, 2026-07] — the shotlist's hardest judgment call is how many script beats share one 15s prompt. There is no fixed ratio (canonical productions ran anywhere from 1.4 rows per prompt to 4.7); decide per scene with this heuristic:

Group shot rows into ONE prompt when ALL of these hold:

1. Same character set in frame
2. Same location (or sub-area of it)
3. One continuous emotional/temporal unit — no time skip, no mood pivot
4. Stageable in ≤15 seconds of screen time
5. The combined prompt stays inside practical length limits (ZH: the 1,800-char hard cap; EN block prompts run much longer — see ../higgsfield-seedance/SKILL.md § Field calibration)

Split into separate prompts when ANY of these fire:

1. Hard cut between locations (apartment → flashback)
2. A major character entrance/exit changes the handle list
3. A lens/setup change that needs its own envelope (wide establish → tight insert)
4. A performance arc that deserves its own 15 seconds — a reaction that builds across 5–7 beats is never bundled with action. Don't fragment grief: one continuous emotional collapse is ONE prompt even if the script writes it as three rows.
5. An insert/cutaway to a prop or screen (those get their own ECU prompt)

Complexity budget per prompt (split triggers): more than 2 strong actions · more than 2 camera moves · more than 3 important characters · more than 1 complex VFX event · more than 1 location change — any of these means the scene wants another envelope. Duration ladder: 4–8s = one strong action · 8–12s = one action + a reveal · 12–15s = 2–3 simple beats · complex fight/chase/transformation = multiple prompts. Reconciling the ladder with the 15s target: 15s is the default envelope, not a straitjacket — when a scene's beats genuinely fill only 4–8s of screen time, generate a deliberately shorter clip (Seedance accepts 4–15s) rather than padding dead air into 15; the prompt-law's "don't pad" rule wins.

When in doubt, err toward more prompts and shorter envelopes — Seedance handles tight prompts better than overloaded ones, and the user can run them in sequence.

Auto-enrichment for thin briefs. When a scene row is thin ("a guy in a room, he's angry"), don't ask — fill in production detail with the default cinematic choices, never details that change the meaning: 16:9 · one clear physical action · slow controlled dolly-in or locked-off frame · a neutral-to-portrait lens (63°/47° FOV in Seedance block prompts — mm vocabulary like "35–50mm" is for non-Seedance surfaces; 29° only if a close-up needs it — ../higgsfield-seedance/SKILL.md § FOV anchors) · motivated practical light · subtle ambience + one meaningful SFX · a clear final frame. Duration is never auto-defaulted — inside a shotlist the envelope law above governs (target 15s, shorter deliberate clips per the ladder); for a standalone prompt, ask — the seedance rule "always ask the user for runtime, never default" stands.