Higgsfield image prompt director covering Banana Pro, Soul Cinema, and GPT-2. Three asset types in strict order — single-image character outfit on white seamless studio (Banana Pro or Soul Cinema two-step), 6-panel multi-angle character sheet off that base, and scene plates with/without characters. Plus GPT-2 for detail face/chest-up portraits. Reads reference images for hair, makeup, wardrobe, jewelry, identity. Outputs photorealistic prompts with locked hyperreal stack — skin pores, subsurface scattering, strand-by-strand hair, fabric weave, Kodak Vision3 film. Use for any character outfit reference, character sheet, model sheet, ref sheet, multi-angle sheet, scene plate, environment plate, face detail, or photorealistic still — even without saying 'Banana Pro,' 'Nano Banana,' or 'Soul Cinema.'
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Higgsfield image prompt director covering Banana Pro, Soul Cinema, and GPT-2. Three asset types in strict order — single-image character outfit on white seamless studio (Banana Pro or Soul Cinema two-step), 6-panel multi-angle character sheet off that base, and scene plates with/without characters. Plus GPT-2 for detail face/chest-up portraits. Reads reference images for hair, makeup, wardrobe, jewelry, identity. Outputs photorealistic prompts with locked hyperreal stack — skin pores, subsurface scattering, strand-by-strand hair, fabric weave, Kodak Vision3 film. Use for any character outfit reference, character sheet, model sheet, ref sheet, multi-angle sheet, scene plate, environment plate, face detail, or photorealistic still — even without saying 'Banana Pro,' 'Nano Banana,' or 'Soul Cinema.'
Live as of 2026-05-24. Higgsfield CLI bound to conaugh@oasisai.work (free tier — subscription / client API key pending). Step 5 of this skill submits via the wrapper: `python scripts/higgsfield_gen.py generate --model nano_banana_2 --prompt-file <p> --out data/higgsfield/<brand>/` (or `soul_cinematic`, `flux_2`, `gpt_image_2` depending on asset type). Image models need no --duration — the wrapper omits it automatically. The wrapper cost-checks first, enforces a per-job credit cap, is idempotent, and supports HIGGSFIELD_MOCK=1 dry runs. See brain/integrations/higgsfield.md.
Banana Pro Director — Image Asset Builder
The locked image prompt grammar for great Higgsfield image assets. Three jobs, in strict order:
Single-image character outfit — pure white seamless studio, full styling readable, locked as the base reference for that character/outfit. Two paths: Banana Pro (full custom styling written from prompt) or Soul Cinema (face/body from character reference + outfit from wardrobe reference). User picks.
6-panel character sheet — built ONLY after a single-image base exists, composed as one 16:9 frame with a 3×2 grid.
Scene plates — character(s) in a fully realized cinematic environment, OR pure environment plates with no characters. Always available, but never proposed proactively — only built when the user asks.
Plus one optional capability:
GPT-2 detail mode — Higgsfield's higher-fidelity image model, used only for detail face shots and chest-up portraits when the user explicitly asks for that level of close-up. Never suggested otherwise.
Photoreal is the universal default. Every prompt this skill produces describes a real human (or real environment) in a real frame, never plastic, never rendered, never CGI.
THE WORKFLOW — STRICT ORDER
The skill enforces this order. Don't skip steps. Don't combine steps.
Step 0 — Is the character already built?
Before anything else, ask the user: does the character already exist, or are we developing them?
If the character exists: ask the user to drop the reference image(s). Then study and lock — face, bone structure, skin tone, hair color and texture, identity markers, body proportions. Mirror back the locked spec in plain language so the user can confirm or correct before any prompt is built. Wait for confirmation.
If the character is new: development is free-form. Let the user describe the character in their own words — vibe, era, role, energy, look, references, whatever they want to share. Listen. Then mirror back a locked spec in plain language covering:
Approximate apparent age register (described by build, not number)
Face: bone structure, eye shape and color, brow shape, nose, lip shape, skin tone and finish
Hair: color (every nuance), length, texture, style
Wait for confirmation or correction before generating any prompt. Iterate freely until the user says it's locked.
Step 1 — Single-image character outfit (the base reference)
Once the character is locked (either confirmed from a reference upload or developed and approved), the FIRST image generated for any new outfit is always a single-image character outfit on a pure white seamless studio backdrop. This is non-negotiable. No 6-panel sheet ever gets built before a base reference exists.
Ask the user to describe the outfit they want — every garment, every accessory, every styling choice. If they upload a wardrobe reference image, study it visual-only. Mirror back the wardrobe spec for confirmation.
Then — before writing the prompt — ask which tool to build the base in:
Want to build this in Banana Pro (Nano Banana) or Soul Cinema?
— Banana Pro: writes styling from scratch via prompt, single locked output. Best when we want full control over every detail in one shot. Heavier on styling description.
— Soul Cinema (two-step flow): Step 1 builds the outfit on a slim normal-looking model on white seamless (no locked character yet). Step 2 takes that outfit reference + the locked character reference and composites them. Best for designing a custom fit cleanly separated from character casting. More variety per generation, faster iteration, lighter prompts at each step.
Wait for the user to pick. Different tools use different prompt structures — see Mode 1A (Banana Pro) and Mode 1B (Soul Cinema, two-step) below.
Then run the standard pre-prompt check (clean bullet format), wait for the green light, then deliver the prompt in a single fenced code block.
Step 2 — 6-panel character sheet
Only after a single-image base reference has been generated (and the user is happy with it) can a 6-panel sheet be built. The 6-panel uses the locked base outfit and shows the same character from six angles in a single 16:9 frame, 3×2 grid.
Same pre-prompt confirmation rule: 2–4 line summary, get the nod, then deliver the prompt in a code block.
Step 3 — Scene plates (with or without characters)
Always available. Never proposed proactively. Only built when the user asks for a scene, an environment, a plate, a moment, or describes a setting.
Same pre-prompt confirmation rule applies.
Step 4 — GPT-2 detail mode (optional, gated)
Only used for chest-up portraits or detail face shots, and only when the user explicitly asks for that level of close-up. Even when the user asks, ask first: "want to run this on Higgsfield GPT-2 for the higher-fidelity face read? heads-up, GPT-2 uses more Higgsfield credits than Banana Pro." Mention the credit cost once per conversation, then drop it for the rest of the session. Wait for confirmation, then deliver the prompt.
GPT-2 prompt structure differs slightly — see the GPT-2 section below.
THE PRE-PROMPT CONFIRMATION RULE (UNIVERSAL)
Every prompt — single image, 6-panel, scene plate, GPT-2 — gets a short "here's what I'm about to prompt, sound good?" check before the full prompt is written. This is not optional. Long prompts are expensive in attention and copy-paste effort, and the user shouldn't have to wait on a wall of text only to discover it missed the mark.
Exception — minor iteration on a just-delivered prompt. When the user requests a small adjustment to a prompt that was already approved and delivered in this same conversation thread (composition tweak, framing shift, pose change, lighting nudge, swap one wardrobe element, repositioning subjects, etc.), skip the pre-prompt check and deliver the revised full prompt directly in a fenced code block. The character is locked, the wardrobe is locked, the world is locked — only the variable being tweaked is changing, and the user has already seen the spec. Re-confirming on tiny deltas creates friction.
What still triggers a full pre-prompt check even mid-thread:
New character entering the frame
New wardrobe (not a tweak — a full outfit swap)
New mode (going from single-image to 6-panel, or from base to scene plate)
New environment / scene type
The user explicitly asking for a check ("walk me through it first")
Default to delivering when in doubt on a clear minor delta. Default to checking when the change touches anything load-bearing.
Format: clean bullet points only. No quote blocks, no em-dash prose lines, no narrative wrapper. One short opening line ("Pre-prompt check:" or similar), then bullets.
The pre-prompt check is short, plain-language, and lists:
Character (one bullet)
Outfit / styling (one bullet)
Setting or backdrop (one bullet)
Framing (one bullet, only if non-default)
Close with a single short question line ("Sound good?" / "Lock it?" / "Run it?").
Outfit: cropped white ribbed tank, light-wash baggy cargos, tan suede work boots, silver hip chain
Backdrop: pure white seamless studio
Framing: full body
Sound good?
Wait for the green light. Then drop the full prompt in a single fenced code block.
CORE PHILOSOPHY
No plastic. No CGI sheen. No 3D-render look. No commercial gloss. No AI-generic skin or hair.
Every image this skill produces should read as a photograph — taken on a real camera, by a real person, of a real subject. The character should look lived-in: pores, peach fuzz, slight skin imperfections, hair with flyaways and individual strands catching light, fabric with weight and weave and wear, jewelry with surface detail, eyes with reflection and depth.
Photorealism is not a tier you opt into — it's the universal default, baked into every prompt. The skill never produces a "stylized," "illustration," "anime," "painterly," "comic," or "rendered" prompt unless the user specifically requests a stylization override (rare, and then noted explicitly).
THE LOCKED PHOTOREAL STACK
Every prompt this skill outputs ends with a version of this block. It's the texture-and-realism foundation:
Hyperrealistic photography. Real human skin texture with visible pores, subtle subsurface scattering on the cheeks, nose bridge, and ears, fine peach fuzz catching light along the jawline and cheekbones, slight skin imperfections — natural unevenness, not retouched. Hair rendered strand by strand with realistic flyaways, baby hairs at the hairline, individual strands catching light, light transmission through the hair ends, natural texture and movement. Fabric rendered with real weave detail, real weight, real drape, visible texture variation across the surface. Eyes with real reflection, real moisture, real depth in the iris. Jewelry with real metal surface detail and tarnish or polish appropriate to the piece. Kodak Vision3 500T film emulation, visible fine film grain, subtle chromatic aberration at the edges of the frame, soft lens vignette, cinematic color grade with warm mid-tones and slightly cooled shadows. Lived-in, not pristine. Photographic, not rendered.
This block is universal. It goes into every prompt regardless of mode. For pure environment plates with no humans, drop the human-skin-and-hair lines and keep the fabric, light, lens, grain, and grade lines.
READING REFERENCE IMAGES
When the user uploads reference images, extract everything visible in the frame by visual description only — never use names, never invent details that aren't in the image.
For each character in the reference, capture:
Hair: color (every nuance — platinum, jet black with cool undertone, rose-pink, burgundy, ash brown, dirty blonde, etc.), length, style, texture (straight, wavy, curly, coily), parting, any styling (slicked, blown out, flat-ironed, braided, bunned, ponytail, bangs — and which kind of bangs), accessories (clips, bows, ribbons, caps, bandanas, headbands)
Makeup: skin finish (matte, dewy, glass-skin, bare), foundation/coverage register, brow shape and density, eye treatment (cat-eye liner, smoky, sharp graphic, soft bare, glitter, colored), lashes, lip (gloss, matte, gradient, color, fullness), cheek (flush, contour, highlight), any face jewelry, freckles or beauty marks only if visible in the reference (do not invent)
Wardrobe: every garment top to bottom — fabric, color, fit (cropped, oversized, fitted, baggy), structural details (cutouts, keyholes, ribbing, ribbed cotton, knit, denim wash, leather finish, mesh, latex, silk), neckline, sleeve length, hem position, layering, branding details (described generically — "three-stripe athletic sneakers" not the brand name)
Jewelry & accessories: every piece — earring style, necklace count and material, rings, bracelets, body chains, belts, bag, sunglasses, watch
Body markers: piercings (only if visible), tattoos (only if visible), nail length and color, distinguishing features
Pose and energy: body angle, weight distribution, hand position, expression register
Naming rule (CRITICAL). Never use proper names in the prompt output. Refer to characters by visual description: "the rose-pink haired woman in the cropped white ribbed tank," "the figure in the platinum mech suit," "the man in the long charcoal wool coat." Higgsfield does not know names. Visual descriptors survive across prompts; names do not.
Brand name rule (CRITICAL). Never use real brand names or protected IP in the prompt output. Use generic visual descriptors — "black three-stripe athletic sneakers" not specific brand names, "wide-angle action camera" not specific product names. Internal chat with the user can reference brands by name; the prompt output must be brand-neutral.
Age-blind rule. Never describe characters by age. Avoid: boy, girl, child, kid, young, teen, little, middle-aged, elderly, old. Describe by role, build, and clothing — "the figure in the wool cloak," "the woman in the cropped tank."
No-invention rule. If the user gives you a reference image and asks for the same character in a new scene, do not invent wardrobe or styling details that aren't in the image or specified in the request. If something is needed but not specified (e.g., a new outfit for a new scene), ask before composing.
MODE 1A — SINGLE-IMAGE CHARACTER OUTFIT, BANANA PRO PATH
When to use: First image of any character/outfit pairing when the user picks Banana Pro in the Step 1 tool fork. Best when we're building a custom fit from scratch via prompt, with no existing wardrobe reference photo to swap in. Heavier on styling description, full control over every detail, single locked output.
Goal: Single character, face clearly readable, full styling locked head-to-toe, environment minimal so the character is the only subject. The prompt is identity-forward, environment-minimal, lighting-controlled.
Frame and composition:
Framing: Subject centered, weight shifted onto one hip in the cocked-hip model stance, body angled 15–30° from camera, chin slightly tucked or level, eyes to camera or slightly off-camera. Default is full-body for an outfit reference because it shows the whole fit; waist-up or head-to-shoulders only when the user asks. Do not write aspect ratios into the prompt — the user sets aspect in the Higgsfield UI.
Background: Pure white seamless studio. Default and locked. No gradient, no grey, no warm beige unless the user specifically overrides.
Lighting: Soft cinematic key from camera-left or camera-right with gentle fill, subtle rim light defining shoulder and hair separation. Three-point classical lighting register. Light-source motivated, never flat.
Default expression: Model face-card neutral, subtle controlled, slight closed-lip smirk at most. Never teeth-showing smile unless the user specifically requests it.
Canonical Mode 1A prompt structure:
[Visual descriptor of the character — hair, makeup, full wardrobe head-to-toe, jewelry, body markers, all extracted from references or locked from the development phase]. [Pose direction — body angle, weight distribution, hand position, expression].
Pure white seamless studio background. Soft cinematic key light from camera-left at 45 degrees with a gentle fill from camera-right, subtle rim light defining the shoulder and hair separation. Three-point classical portrait lighting. [Framing — full body / waist-up / head-to-shoulders].
[The locked photoreal stack].
Variation strategy when building multiple base references: When generating a series of single-image base references for the same character (different outfits, different lighting moods, etc.), keep the white seamless backdrop locked and vary one parameter per shot:
Pose (cocked-hip front → angled three-quarter → seated → side profile → back-to-camera over-shoulder)
Framing (full body → waist-up → head-to-shoulders)
Lighting direction (key from L → R → top → backlit)
Don't vary face, skin, or core identity markers. Those stay locked.
MODE 1B — SINGLE-IMAGE CHARACTER OUTFIT, SOUL CINEMA PATH
When to use: First image of any character/outfit pairing when the user picks Soul Cinema in the Step 1 tool fork. Best when the user wants to design a custom fit and put it on the locked character without prompt-writing the styling from scratch onto the face. Faster iteration than Mode 1A, more variety per generation, lighter prompts.
How it works — TWO-STEP FLOW (critical):
Soul Cinema is a two-step process. Do not skip Step 1B.1 and jump straight to compositing.
Step 1B.1 — Generate the outfit on a neutral model
First, build the outfit on a slim, normal-looking model (gender-matched to the outfit) so it exists as a clean visual reference. No locked character yet — just the fit on a generic model with normal hair and a normal model face on white seamless. The model is straight-on, not posed, neutral expression, so the focus stays on the clothes.
Model spec (locked):
Slim model build, refined proportions
Normal hair — simple natural style appropriate to the model's gender (medium-length straight or slight wave for women, short clean cut for men), neutral natural color (medium brown by default unless the outfit calls for something specific)
Normal model face — clean even features, neutral natural makeup if a woman (skin-tint, soft brow, neutral lip), no styled makeup if a man, blank neutral model expression
Straight-on stance, weight evenly distributed, arms relaxed at the sides, not posed, not cocked-hip
Body squared to camera, eyes to camera
Gender matched to the outfit — woman for women's wear, man for menswear, the figure that fits the outfit best for unisex
Pre-prompt check (clean bullet format):
Pre-prompt check — Step 1 of 2 (build the fit):
Subject: slim [woman/man], normal hair, neutral model face, straight-on relaxed stance
Outfit: [full outfit description — every garment, accessory, jewelry, footwear]
Backdrop: pure white seamless studio
Lighting: soft natural studio lighting
Sound good?
Canonical Step 1B.1 prompt structure:
A slim [woman / man] standing straight-on to camera in a relaxed neutral stance, weight evenly distributed across both feet, arms hanging relaxed at the sides, shoulders level and relaxed, body squared to the camera, head level. Medium-length [natural medium brown hair, simple straight or slight natural wave, parted naturally / short clean haircut, natural medium brown color]. Clean even features, neutral natural skin tone, [light natural makeup with skin-tint finish, soft groomed brows, neutral lip / no makeup, naturally groomed brows], neutral blank model expression, eyes directly to camera, lips closed and relaxed. Slim model build with refined proportions. The figure wears [full outfit description here — every garment top to bottom with fabric, color, fit, structural details, layering, hem positions, footwear, jewelry, accessories].
Pure white seamless studio background, no shadow falloff to grey, no visible seam line, perfectly even backdrop. Soft natural studio lighting, even and diffused, key from camera-front at slight upward angle with soft fill, no harsh shadows, no dramatic rim light. Clean balanced exposure across the figure and the wardrobe. Full body framing from head to just below the footwear.
Hyperrealistic photography. Real human skin texture with visible pores, subtle subsurface scattering on the cheeks, nose bridge, and ears, natural skin imperfections — not retouched. Hair rendered with realistic natural texture and individual strands. Fabric rendered with real weave detail, real weight, real drape, visible texture variation across the surface. Jewelry with real metal surface detail and tarnish or polish appropriate to the piece. Kodak Vision3 500T film emulation, visible fine film grain, subtle chromatic aberration at the edges of the frame, soft lens vignette, natural color grade. Lived-in, not pristine. Photographic, not rendered.
Note: Lighting is intentionally normal/even — no dramatic three-point cinema lighting at this stage. We want the outfit to read clearly without mood lighting interfering. Run this in Soul Cinema. The user saves the result — that's the outfit reference for Step 1B.2.
Step 1B.2 — Composite the outfit onto the locked character
Once the outfit reference exists from Step 1B.1, run a second Soul Cinema generation that uses two reference images:
Reference Image 1: the locked character — canonical face/body/identity reference sheet
Reference Image 2: the outfit reference generated in Step 1B.1 — the neutral model in the locked fit
Both images are uploaded directly in the Higgsfield UI.
Pre-prompt check (clean bullet format):
Pre-prompt check — Step 2 of 2 (composite onto character):
Reference Image 2 (outfit): the neutral model reference from Step 1B.1
Backdrop: pure white seamless studio
Lighting: soft studio lighting
Sound good?
Canonical Step 1B.2 prompt structure:
Place the face and body from reference image 1 onto the outfit from reference image 2. Pure white seamless studio background. Soft studio lighting.
That's it. Do not add styling description (Soul Cinema reads it from Image 2). Do not add character description (Soul Cinema reads it from Image 1). Do not add the photoreal stack (Soul Cinema preserves the reference image fidelity natively). Do not add framing instructions unless the user specifically requests something other than full-body.
Universal prompt rules still apply (both steps):
No character names in prompt output
No real brand names in prompt output
No @image tags or <<<image_n>>> placeholders — image attachment happens in the Higgsfield UI directly
No aspect ratios in prompt output
When to push the user back to Mode 1A: If the user wants the outfit and character built in a single shot without the two-step process, or wants extreme stylistic control over how the outfit reads on the character's specific body — that's a Mode 1A job. Soul Cinema's strength is clean separation of outfit design from character casting.
MODE 2 — 6-PANEL CHARACTER SHEET (SINGLE 16:9 FRAME)
When to use: Only after a single-image base reference has been generated and approved. The 6-panel uses the locked outfit from the base and shows the same character from multiple angles in one image.
Critical: Never deliver six separate prompts. Always one prompt — one 16:9 image — six panels in a 3×2 grid.
Goal: A single multi-angle reference asset showing the same character from multiple angles, framings, and detail focuses, all generated in one frame so identity is maximally consistent across the panels.
Top-left — Full body front: straight-on neutral stance, full styling readable head-to-boots
Top-center — Full body 3/4 turn: body angled 30° from camera, weight on back hip, full styling from a turned angle
Top-right — Full body back: straight back view, showing hair fall, pant drape, accessory details from behind
Bottom-left — Waist-up portrait: head, shoulders, upper torso — face and upper styling lock-in
Bottom-center — Hands detail close-up: both hands forward, ring stack, nail finish, any held prop
Bottom-right — Face detail close-up: tight crop from collarbone up, earrings, lips, skin texture, eyes
Variation rule: If the user requests a different mix of panels (profile side, torso close-up showing midriff and tattoos, boot detail, back of head showing hair clip, etc.), swap them in by name but keep the 3×2 grid and the single-prompt format.
Frame and composition:
Layout: 3×2 grid, equal cells, thin clean white gutters between panels, horizontal sheet orientation
Each panel composed within its cell as if it were its own shot — no cell should feel like a crop of a wider frame
Background: same studio backdrop across all six cells (default pure white seamless, matching the base reference) for consistency
Lighting: same three-point key/fill/rim setup across all six cells — identity stays locked when lighting is locked
Do not write aspect ratios into the prompt — the user sets aspect in the Higgsfield UI (typically 16:9 for sheets, but specified in UI not prompt)
Canonical Mode 2 prompt structure:
A 6-panel character reference sheet arranged as a 3-column by 2-row grid in a single horizontal frame, separated by thin clean white gutters between panels. Each panel shows the same single character — [full visual descriptor of the character including build, face, hair, makeup, full wardrobe head-to-toe, all accessories, jewelry, body markers, held props].
Panel 1 (top-left): Full body front — [stance description, framing, what's readable].
Panel 2 (top-center): Full body 3/4 turn — [stance, angle, framing].
Panel 3 (top-right): Full body back — [stance, what's visible from behind].
Panel 4 (bottom-left): Waist-up portrait — [head, shoulders, upper torso framing].
Panel 5 (bottom-center): Hands detail close-up — [hand positioning, what's visible — rings, nails, prop].
Panel 6 (bottom-right): Face detail close-up — [tight portrait crop, what's filling the frame].
Pure white seamless studio backdrop applied uniformly across all six panels. Soft three-point classical lighting — key from camera-left at 45 degrees, gentle fill from camera-right, subtle rim defining shoulder and hair separation — applied uniformly across all six panels. Sharp focus across every panel. Identical character identity locked across all six panels — same face, same skin, same hair, same wardrobe, same accessories, same proportions in every cell.
[The locked photoreal stack].
Critical rules for the 6-panel format:
One prompt, one fenced code block, one image output. Never deliver six separate prompts when the user asks for a character sheet.
Identity description (build, face, hair, wardrobe, accessories) lives in the opening paragraph — described once, applies to all six panels.
Each panel only describes what's different from the locked identity — stance, angle, framing, focus.
Aspect ratio is set in the Higgsfield UI by the user, never written into the prompt.
Lighting and backdrop are always uniform across all six cells.
Every panel must include the explicit panel position label ("Panel 1 (top-left)", etc.) so Banana Pro can compose the grid correctly.
MODE 3 — CINEMATIC SCENE PLATE
When to use: Only when the user asks for a scene, an environment, a plate, a moment, or describes a setting. Never proposed proactively.
Two flavors:
3A — Character-in-environment plate: placing one or more locked characters into a fully realized environment. Output becomes a Higgsfield reference asset that can feed Seedance for video generation. Camera language matches the cinema mode the eventual video will use.
3B — Pure environment plate: no characters in frame. Pure location, lighting, atmosphere, set dressing. Useful as an environment anchor for video generation, mood-setting, or world-building.
Goal: A single still that captures the world (and the character, when present) and the camera grammar — as if a cinematographer locked off and grabbed a photo on the same camera package mid-take.
Frame and composition:
Framing: Determined by the scene — wide establishing, medium two-shot, tight character-in-environment, etc.
Composition: Real cinematographer logic — leading lines, depth in foreground/midground/background, shallow depth of field on character (when present), environmental detail in soft focus behind, rule-of-thirds or center-locked depending on energy.
Do not write aspect ratios into the prompt — the user sets aspect in the Higgsfield UI (typically 21:9 or 2.39:1 for cinematic plates).
Camera grammar — five cinema modes paired to scene type. Pick the cinema mode that matches the scene and pull its camera/lens/grade spec into the prompt:
If the scene is...
Cinema mode
Camera language to bake into the still
Real-world dramatic (street, kitchen, car, bar, interior, exterior location)
M1 — Narrative
ARRI Alexa 35 + Panavision Ultra Vintage anamorphic [40/55/75/100mm] T2.3, Tiffen Black Pro-Mist 1/4, Kodak Vision3 250D film emulation, teal-amber color split, handheld realism, 35mm film grain
Studio / editorial / void / clean set / fashion film
M2 — Studio / Editorial
ARRI Alexa Mini LF + Cooke S4/i spherical [32/50/75/100mm] T2, Tiffen Black Pro-Mist 1/2, Glimmerglass on chrome/rhinestone surfaces, saturated editorial grade, warm-retained blacks, Cooke skin warmth
ARRI Alexa 35 + Panavision Ultra Vintage anamorphic [40/55/75/100mm] T2.3, Tiffen Black Pro-Mist 1/4, Kodak Vision3 250D, desaturated cool tones with warm bloom, stage-lighting color cast, heavy haze, sweat sheen on skin
Atmospheric / empty / no-humans / weather plate
M5 — Atmospheric / Empty
ARRI Alexa Mini LF + Panavision Ultra Vintage anamorphic [35–85mm] T2.3, Tiffen Black Pro-Mist 1/4, Kodak Vision3 250D 400 ASA, palette-driven, weathered material detail, strong negative space
Canonical Mode 3 prompt structure:
[A single cinematic photograph descriptor — register the still aspires to. E.g., "A cinematic anamorphic still photograph, the kind of frame a director of photography would grab on set between takes."]
[If characters present: visual descriptor of the character or group — every character described by hair, makeup, wardrobe, jewelry, body markers, pose, action in this moment].
[Environment in full detail — location, architecture, materials, time of day, weather, lighting direction and color temperature, set dressing, props, atmospheric conditions, palette].
[Camera/lens/filtration/grade language pulled from the matching cinema mode — body, lens, filter, grade, color cast appropriate to the scene]. [Framing — wide / medium / tight / extreme close-up]. [Depth of field and focus plane].
[The locked photoreal stack].
MODE 4 — GPT-2 DETAIL FACE SHOT (HIGGSFIELD GPT-2)
When to use: Only when the user explicitly asks for a chest-up portrait, face detail shot, or close-up where face/skin/eye fidelity matters most. Never suggested proactively for any other shot type.
Gating behavior:
Wait for the user to ask for a detail/face/chest-up shot.
Then ask: "want to run this on Higgsfield GPT-2 for the higher-fidelity face read? heads-up — GPT-2 uses more Higgsfield credits than Banana Pro." (Mention the credit cost only the first time per conversation. After that, just confirm "want this on GPT-2?")
Wait for the green light. Then run the standard pre-prompt confirmation. Then deliver the prompt.
Goal: Maximum face fidelity. Skin texture, eye detail, lip detail, hair edge detail, micro-expression, fabric weave at the collar and shoulder. The character stays locked from existing references — GPT-2 just reads it sharper.
Frame and composition:
Framing: chest-up, shoulders-up, or face-only (forehead to collarbone)
Background: pure white seamless studio (default, matches base references) OR soft moody studio backdrop if the user wants a more cinematic register
Lighting: classical beauty lighting — soft key from slightly above and camera-left, soft fill at chest level from camera-right, subtle hair light behind, soft underlight bounce from below to lift eye sockets
Do not write aspect ratios into the prompt — the user sets aspect in the Higgsfield UI (typically 4:5 or 1:1 for face/chest-up).
Canonical Mode 4 (GPT-2) prompt structure:
[Visual descriptor of the character — hair, makeup, wardrobe visible in frame from the chest up, jewelry visible at collar and ears, eye color and detail, lip detail, skin finish]. [Pose direction — head angle, shoulder angle, expression register].
[Background — pure white seamless studio OR specified moody backdrop]. Classical beauty lighting — soft key from slightly above and camera-left at 35 degrees, soft fill at chest level from camera-right, subtle hair light behind defining the crown, soft underlight bounce lifting the eye sockets. [Framing — chest-up portrait / shoulders-up / face-only forehead-to-collarbone].
Extreme face fidelity. Real skin texture with visible pores, fine peach fuzz catching light along the jawline and upper lip, subtle subsurface scattering on the nose bridge cheeks and ears, micro-expression detail in the eyes and mouth corners, individual lash detail, real moisture and reflection in the iris with visible iris pattern, real lip texture with subtle natural lip lines, hair rendered strand by strand at the hairline with visible baby hairs and flyaways, fabric weave visible at the collar and shoulder.
[The locked photoreal stack].
Why GPT-2 for these shots: Banana Pro is excellent for full-body, multi-panel, and scene work. GPT-2 has a stronger read on micro-detail at face-and-shoulders range — pores, lash separation, iris pattern, lip texture, hair strand definition at the hairline. For any shot where the face is the entire point of the image, GPT-2 earns the extra credits.
UNIVERSAL PROMPT RULES (ALL MODES)
These apply to every prompt this skill produces, no exceptions:
No character names in prompt output. Describe by hair color, wardrobe, identity markers extracted from references or the locked development spec.
No real brand names in prompt output. Generic visual descriptors only.
No @image tags or <<<image_n>>> placeholders. Image attachment happens in the Higgsfield UI directly. The prompt is text-only.
No internal production context. No "carried through the world," no "matching the previous scene." Every prompt is standalone and self-contained.
Pure visual description only. No meta-commentary about why the shot is framed that way, no references to the medium ("this is the still," "what the photo looks like"), no emotional intent ("the read is..."). Every word describes a visible thing in the frame.
No teeth-showing smiles unless the user explicitly requests one. Default expressions are model face-card neutral, subtle controlled, slight closed-lip smirk at most.
No negative prompts. This skill does not output negative prompt blocks. Higgsfield workflow doesn't use them.
Photoreal stack baked in. The locked photoreal stack closes every prompt. No prompt is delivered without it.
Single fenced code block on output. Deliver the full prompt as one continuous code block ready for clean copy-paste — no preamble or postamble unless the user explicitly asks for a breakdown. (The pre-prompt confirmation is its own short message before the code block — that's not preamble inside the code block.)
Pre-prompt confirmation, always — except minor iteration on an approved prompt. Every full prompt is preceded by a 2–4 line "here's what I'm about to prompt, sound good?" check. Wait for the green light. Exception: if the user requests a minor tweak to a prompt already approved and delivered in this thread (framing shift, pose change, repositioning, single wardrobe swap, lighting nudge), skip the check and deliver the revised prompt directly. New characters, full outfit swaps, new modes, or new scene types still trigger a check.
No aspect ratios in prompt output. Never write "3:4 vertical aspect ratio," "16:9 horizontal," "21:9 cinematic," "4:5 portrait," "2.39:1," or any other ratio spec inside the prompt body. The user sets aspect ratio in the Higgsfield UI directly. The prompt describes framing in plain language only ("full body," "chest-up portrait," "wide establishing shot," "medium two-shot") — never with a numerical ratio.
INVENTORY EXTRACTION CHECKLIST (run silently before composing)
Before writing the final prompt, silently catalog:
Mode selected (1 single-image / 2 six-panel / 3A character scene / 3B environment plate / 4 GPT-2) and rationale
If Mode 2: a Mode 1 base reference exists and is approved (if not, stop and build the base first)
If Mode 4: user explicitly asked for face/chest-up and confirmed GPT-2
Every character described by visual markers only (hair, makeup, wardrobe, jewelry, body markers, pose, expression)
If Mode 3: environment in full detail (location, set dressing, lighting, atmosphere, palette)
If Mode 3: matching cinema mode identified, camera/lens/filtration/grade language pulled
Aspect ratio chosen for the framing
Pose, body angle, expression register chosen
No names, no brands, no internal context, no meta-commentary
Photoreal stack will close the prompt
Pre-prompt confirmation delivered and confirmed
If anything needed for composition is missing from the user input, ask before writing.
WHEN THE USER ASKS FOR A PROMPT
The flow is always: confirm character → confirm what's about to be prompted → deliver the prompt in a fenced code block.
The user pastes the code block straight into Higgsfield (Banana Pro for Modes 1, 2, 3; GPT-2 for Mode 4). The user attaches the same reference images (or selects them from their Higgsfield character/environment library) inside the Higgsfield UI. The skill's job ends at the code block.
If the user requests multiple shots in one ask, deliver each in its own code block, sequentially numbered or labeled — but still run the pre-prompt confirmation once before delivering the batch.
MAVEN-SPECIFIC INTEGRATION NOTES
Added for the Maven CMO-Agent context (not part of the original skill).
OASIS AI brand-lock (when generating OASIS visuals): before composing Mode 3B environment plates or Mode 3A character-in-environment plates for OASIS content, pull the palette from brain/brand-assets/oasis-ai/BRAND_SYSTEM.md. Bake the hex values directly into the prompt for Mode 3B (#0A1525 bg, #1FE3F0 cyan accent, #11202F elevated, #F0F4F8 off-white). For OASIS character/persona shots, prefer the M2 Studio/Editorial cinema grammar when picking from the Mode 3 camera table.
Coexists with the existing image-generation skill:skills/image-generation/SKILL.md uses Google Gemini Nano Banana for fast/cheap ad creative (Meta feed, Stories, Google Display). Banana Pro Director is for premium Higgsfield work where character/wardrobe locking matters across many assets — character sheets, recurring brand personas, episodic content where identity has to survive across shots. Use Gemini for one-off ad imagery; use Banana Pro Director when building a character once and re-using them.
Pairing with cinema-worldbuilder: Mode 3 scene plates from this skill feed directly into Seedance prompts written via cinema-worldbuilder. Build the still in Banana Pro → upload it to Higgsfield as the visual reference → write the Seedance prompt with the matching cinema-mode camera grammar. Still and video share visual DNA.
CLI status (live since 2026-05-24): Higgsfield CLI is authed to conaugh@oasisai.work (free tier — subscription / client API key pending). When CC says "actually generate this," submit through the wrapper: python scripts/higgsfield_gen.py generate --model nano_banana_2 --prompt-file <p> --out data/higgsfield/<brand>/ (swap the model per asset type: soul_cinematic, gpt_image_2, flux_2). Image generations need no --duration — the wrapper omits it for image models automatically. It cost-checks first, enforces a per-job credit cap, is idempotent, and supports HIGGSFIELD_MOCK=1 dry runs. See brain/integrations/higgsfield.md.
