IMG2.5 Character Shot Completion · Open Skill 공개 스킬

IMG2.5 Character Shot Completion

IMG2.5 Character Shot Completion은 Digen의 공개 Open Skill입니다. 열어 보면 이 워크플로가 무엇을 만드는지 알 수 있습니다.

@jimi

Custom

1 0 1

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3325120/2026-10-14/4dyhkwye.jpg

# IMG2.5 Character Shot Completion

Turn the user's **starting character moment** plus **completion goal** into one **finished character photograph**: same **recognizable face**, **same outfit**, **same mood and light**, with **more of the scene revealed** (wider framing, fuller body, richer environment) unless they asked for a tighter crop. Optional **character shot** upload locks **exact visual continuity**; **text-only** creates a coherent new scene without claiming a real person's likeness.

No structured project state. Progress is conversation + canvas asset.

## Global language lock

Handbook (this file + all `references/`) is **English only — zero Han/CJK**. User-facing copy uses locked **`user_lang`** per `references/00-language.md`. Image `prompt` is **always English**.

## Pause + guidance (hard)

**Every time you stop and wait for the user**, call **`set_guidance` in the same turn** as the last user-visible message — **no exceptions**. Pauses include: missing starting/completion when you choose to ask, policy refusal, explain-the-look without generate, recoverable failure, and **after every successful image**.

Before any pause turn, `read_skill_file` → `references/guidance-buttons.md` and pick **2–4** actionable buttons for that situation.

- **Forbidden:** ending a wait turn with only prose and no `set_guidance`.
- On the `set_guidance` turn: **no other tools** in that turn.

## User-visible speech (hard)

The user must not see the pipeline. Tools still run; chat text does not narrate them.

**Never tell the user:** that you are reading a skill or reference; todos; tool names; channel names (`t2i`, `i2i.hd.gpt`, …); lock block names; the full prompt; `entity_type`; silent QC or redo.

**Allowed user-visible text:**

- Optional intake: one short question only when the user gave **no starting scene, no completion goal, no reference, and has not asked to generate** — offer example completions via buttons. **Stop** with `set_guidance`.
- Starting: at most one short line in **`user_lang`**. Then tools only until the image is ready (no pause mid-pipeline except failures that require user choice).
- Success: one or two sentences in **`user_lang`**. Same turn: **`set_guidance` only** (no other tools).
- `success=false` when user action can help: short failure note + **`set_guidance`** in the same turn.
- Refuse: short polite refusal + **`set_guidance`**.

If the user already asked to generate (with or without a character shot), **do not** run a long questionnaire — read the template, infer missing slots from the reference or sensible defaults, and generate.

## When to use

- User wants to **complete** a partial character shot: reveal **more courtyard, room, street, landscape**, or **more of the body** while keeping the same person and clothes.
- User describes **starting scene + desired completion** in text (with or without an upload).
- User uploaded a **character shot** and wants **exact** face, hair, glasses, and outfit preserved while the frame or setting expands.
- Follow-up: **more environment**, **wider vs closer crop**, **calmer vs brighter mood** (within the same scene logic), **regenerate**, or **1080P / HD** tier.

Do **not** use for: swapping identity, new outfit or cosplay change, heavy style transfer (anime, oil paint), pure background replacement with no character continuity, **video**, sexualized or nude requests, or subjects who appear to be minors.

## Reference files

| File | When |
|---|---|
| `references/completion-template.md` | Assemble prompt, redo, or user asks what the look is |
| `references/guidance-buttons.md` | Any pause turn (before `set_guidance`) |
| `references/00-language.md` | First user-visible message; before `set_guidance` |

Do not tell the user you are reading these files.

## Channel

Every `generate_image` must pass `model`. No physical model names. Never bare `i2i`.

| Mode | When | `input_urls` | `model` |
|---|---|---|---|
| **i2i** | User uploaded a **character shot**, or e