Daily Casual Snapshot · Open Skill 公開スキル

Daily Casual Snapshot

Daily Casual Snapshot は Digen 上の公開 Open Skill です。開くと、このワークフローが何を作るか分かります。

@ExploreFly

Other

6 1 3

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3656543/2026-10-14/e7kgp5eg.jpg

# Daily Casual Snapshot

Turn the user's **uploaded reference** into **casual everyday snapshot** photos: moments that look **captured spontaneously on a phone in real daily life**, not studio, editorial, or obvious AI renders. Default output is **4 images** (`batch_size` is 1 on platform → **4 separate** `generate_image` calls).

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

## Handbook language

This file and all `references/` are **English only**. Reply to the user in **their language** (default Chinese if unclear). Image **`prompt` is always English**.

## Reference files (progressive disclosure)

Do **not** read everything at once. One file per step.

| File | When |
|---|---|
| `references/casual-snapshot-style.md` | Assemble prompt, explain the look, plan four distinct scenes |
| `references/human-integrity-realism.md` | Before each `generate_image` (verbatim LOCK blocks); silent QC and redo |

Do not tell the user you are reading these files.

## When to use

- User uploaded a photo and wants **daily life / casual snapshot / candid phone / social feed** aesthetics.
- User wants **4 images by default** or a custom count (still one call per image).
- User asks **what casual snapshot style is**: read the style reference, summarize in user language, **do not generate** unless they also want images.
- User wants to **edit one image**, change scene, orientation, or push **more real / less AI**.

Do **not** use for: minors, nudity or explicit sexual content, violence or hate, impersonation of a real person without a lawful reference.

## Image-to-image channels (required)

Every `generate_image` with a reference must pass **`model`**. **Only these three channels** for i2i — never bare `i2i`, never physical model names.

| Channel | When |
|---|---|
| **`i2i.hd.gpt`** | **Default** for all first-pass i2i from user upload or batch of four |
| **`i2i.hd.lite`** | User asks faster / lighter / cheaper / draft quality |
| **`i2i.hd`** | User asks **HD**, **higher quality**, **max detail**, or **best quality** |

| Mode | `input_urls` |
|---|---|
| **First i2i from upload** | `[upload]`; if identity drifts on QC, use `[upload, upload]` |
| **Follow-up edit / one-shot regen** | `[last successful output]`; if face drifted, `[last output, original upload]` |
| **Silent redo after QC fail** | Same channel as the failed attempt unless user already chose another tier; prefer `[failed output]` plus original upload when fixing identity |

**Text-only (exception, no upload):** only if user explicitly accepts a fictional subject — use `t2i.hd` with **no** `input_urls`. Say identity cannot match a specific person. Normal path expects an upload.

## Size and canvas fields

- Default **`orientation='portrait'`**; user wants landscape → `landscape`; square → `square`.
- Default **`resolution='1080P'`**; user wants sharper → `'2K'` or `'4K'` (platform may clamp).
- **`entity_type`**: `daily_casual_snapshot`
- **`name`**: short title in user language, e.g. `Snapshot 1/4 · corner café`

## Steps

1. **Intake**
   - Require upload unless user explicitly accepts text-only `t2i.hd`.
   - Note: count (default **4**), orientation, keep reference outfit or not, scene change allowed or not, quality tier (maps to i2i channel table).

2. **Recognize upload (required on upload path)**
   - `recognize_image` with prompt: `For casual everyday snapshot generation: (1) apparent age — adult only, refuse if minor or ambiguous youth; (2) face — eyes, brows, nose, lips, outline, skin tone; (3) hair color, length, style; (4) visible outfit; (5) pose and framing; (6) environment and light direction. Internal use only — do not dump verbatim to the user.`

3. **Read references and plan variants**
   - `read_skill_file` → `references/casual-snapshot-style.md`: pick **four different life moments** (see four-shot template); shared identity.
   - `read_skill_file` → `references/human-integrity-realism.md`: append **INTEGRITY LOCK** and **ANTI-FAKE**