---
name: casting
description: "Create a Chinese screen-role casting package from a six-field role card: one pure-white, face-first casting portrait for approval, then only selected identity-bound full-body, expression-grid, or turnaround assets. Use for short-drama, film, and casting-character roles; unlike `create-character` (relationship 0.84), choose this for actor-style face selection and a confirmation gate, and choose `create-character` for a canonical multi-panel production pack and scenes. Do not use for real-person edits, game-role NPC concepts, or generic character sheets."
---

# Casting

## Catalog

- Display name: Casting
- Categories: Character
- Status: Ready
- Featured: yes
- Cover image: /skill-covers/casting.png
- Route label: GPT Image 2 face-first casting package
- Tone: violet
- Short description: Select and approve a screen-ready Chinese character face before deriving optional studio assets.

- Similar skills: create-character, create-npc
## What this skill does

- Converts a role card or script description into a Chinese screen character.
- Generates exactly one vertical, pure-white, front-facing casting portrait first.
- Stops for explicit face approval. Only then creates user-selected full-body, 3×3 expression-grid, and/or turnaround assets.
- Uses separately controlled high-end CG rendering for 漫剧, 3D漫剧, 3D角色, 三维, CG, or 三渲二 requests.

## How to use

Request the six role-card fields when available: person type, build, facial features, skin/hair/makeup, temperament, and intended character type. A script or freeform description is acceptable. Complete only reasonable omissions and state them in one sentence before submission.

Do not reproduce a protected character or edit a real person. References must be authorized. Bind only an accepted public image URL, in supplied order—never invent a URL or asset token. Clothing must not override the requested face, skin, expression, or bone-structure intent.

#### Example

```text
为民国悬疑短剧选一个26岁女记者：身形清瘦，偏窄长脸、平直微挑眉、长眼微上扬，暖白细腻肤色和低盘发，表面温柔但在隐藏锋芒。先出纯白底定脸；我确认后再选全身或表情。
```

## User-facing output

- Media type: Approval-gated image casting package
- Default quantity: One face-setting portrait
- Optional quantity: One selected full-body, expression grid, and/or turnaround board
- Content: Chinese screen-role character on a pure-white background
- Default layout: Vertical 3:4 portrait; selected 3:4 derivatives; 16:9 turnaround board
- Model policy: GPT Image 2 Medium/2K after current-schema verification
- Downstream use: Short-drama, film, comic, and story production casting references

## Route and execution

Use `gpt-image` / GPT Image 2 at Medium/2K. Before a real run, inspect the current official CLI or native Agent schema: the local catalog marks identity consistency and multi-reference support unknown. If public image references are unsupported, stop before derivatives and explain that safe identity continuity cannot be claimed.

## Portrait and approval workflow

1. Default to polished photographic realism. Trigger CG mode immediately only for the listed CG terms; never mix photographic and CG surface wording.
2. For multiple roles, choose one or two differentiating facial/temperament anchors per role. For a single role, follow the brief instead of converging on a generic influencer face.
3. Read [portrait prompt blocks](references/portrait-prompt-blocks.md). Compile one Chinese-led prompt in this order: CHARACTER, SKIN, OUTFIT, then the three locked blocks unchanged. Use CG replacements only in CG mode.
4. Submit one GPT Image 2 task: `quality: "medium"`, `imageSize: "2K"`, vertical `3:4`, `batchCount: 1`, and a persisted operation key. Poll its accepted execution to terminal state; do not replace an unknown create outcome.
5. Inspect before presentation: seamless `#ffffff`; front-facing head-and-neck crop; balanced facial structure; refined natural skin or intended CG surface; clear ears/hair; logo-free clothing; exactly one character.
6. Ask whether the face is approved. If it is not, alter only the failed constraints and submit one linked revision. Never create a derivative before explicit approval.

## Approved expansion

Offer **full-body portrait**, **3×3 expression grid**, and **turnaround board** only after approval. Explain that turnaround needs full body first; do not prepare operation keys or submit unselected assets.

- **Full body:** bind the approved portrait URL in both documented reference fields; use vertical `3:4`, head-to-toe framing, and the applicable written head-to-body and negative constraints from the reference. Women default to seven heads; ordinary men to eight; dominant male leads only may use 8.5–9; antagonists remain realistic; children use 3–4.
- **Expression grid:** bind the approved portrait; use one vertical `3:4` 3×3 grid. Default sequence: neutral, restrained smile, anger, sadness, joy, tearful grief, surprise, intense anger, confusion. Keep orientation, identity, hair, and skin consistent.
- **Turnaround:** bind approved portrait and full-body URLs in that order. Use a 16:9 board with front/profile/back full bodies plus front face close-up. After one material revised attempt, report any layout limitation and offer separate views rather than fabricating consistency.

For each derivative use a new operation key and `batchCount: 1`; reference URLs must match in `input.images` and `params.images`. Poll each accepted execution. A known terminal or visual failure may receive one linked, materially revised retry; reconcile unknown results by the existing key.

## Visual invariants

- Every asset uses seamless pure white `#ffffff`: no gray, off-white, colored, textured, gradient background.
- Photographic skin is clean, refined, realistic, and naturally textured—neither rough/enlarged pores nor plastic glass skin—under soft volumetric face light.
- Describe the fictional subject as `Chinese`, and use state-based temperament rather than stereotypes.
- Keep apparel logo-free and restrained. Tops remain loose over the waistband; default male trousers are full-length straight or gently wide-leg, unless the card requires otherwise.
- Do not alter the locked composition, rendering, and sharpness blocks or add the known harmful focus phrases listed in the reference.

## Acceptance

- Report role, reference bindings, rendering mode, operation key, execution ID, and terminal state; report only successful assets as delivered.
- Pass only when requested original role identity, white background, framing/proportion, no logos, coherent hair/skin/wardrobe, and the approved face remain observable. State model layout/reference limitations plainly.