Amateur Fail Photos-pro · Open Skill 公開技能

Amateur Fail Photos-pro

Amateur Fail Photos-pro 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。

@skyyyyyy

Meme

32 3 9

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3656250/2026-09-22/t5t6vzub.jpg

---
name: amateur_fail_photos
display_name: Amateur Fail Photos
description: Use when the user uploads a character reference and wants a strict match of that image's art style plus face/hair/outfit as amateur fail-photo grids (3×3, 9:16 portrait), cell splits, ref2v micro-clips (≤3s each), or a mix_clips compilation. Also use for redo-one-cell, split-only, animate-some-cells, or restitch-existing-clips. Trigger terms include 素人の失敗写真, fail photos, 3x3, 9:16, split, ref2v, compilation, mix, and "same person / same style as the reference".
allowed-tools:
- generate_image
- split_image
- generate_video
- mix_clips
---

# Amateur Fail Photos → Split → Animate → Stitch

From the user's reference image, generate a **3×3 amateur fail-photo grid**. **Default orientation is portrait 9:16**; use `landscape` (16:9) only if the user explicitly asks. Grid prompt: **Japanese catchphrase → orientation layout block (portrait vs landscape) → English style/person lock**. Do not over-specify what each cell depicts. After split, once the user chooses **Generate video**, run **animate + stitch** (i2v **≤3s per clip** → `mix_clips`) with no extra confirmation.

## When to use

- Amateur fail photos, 3×3 grid, 9:16 portrait, split, ref2v, compilation
- Prompt close to: `図中のキャラクターをモデルにして、素人の失敗写真の数々、3x3`
- **No reference image**: ask for upload first. Text-only fallback is `t2i.hd.gpt` (warn weaker identity match)

## Stage routing (read one file)

| User intent / existing assets | Read |
|---|---|
| Need a grid / just uploaded a reference | `references/stage-grid.md` |
| Full grid exists, user wants split | `references/stage-split.md` |
| User confirms Generate video (cells specified or not) | Ensure split exists, then read `stage-video.md` → `stage-mix.md` **without pausing** |
| ≥2 clips exist and user only wants restitch | `references/stage-mix.md` |

**Do not read all references at once.** Use `read_skill_file` for the current stage only. On the Generate-video pipeline, read video then mix back-to-back — **no `set_guidance` in between**.

## Stages (references/)

| Stage | File | Tool | entity_type |
|---|---|---|---|
| 1. Grid | `references/stage-grid.md` | `generate_image` | `storyboard_images` |
| 2. Split | `references/stage-split.md` | `split_image` | `storyboard_images` |
| 3. Animate | `references/stage-video.md` | `generate_video` | `storyboard_videos` |
| 4. Stitch | `references/stage-mix.md` | `mix_clips` | `final_videos` |

### Cell numbering (all stages)

3×3 left-to-right, top-to-bottom, **1–9**:

```
1  2  3
4  5  6
7  8  9
```

"Center" = 5; "top-left" = 1; "bottom row" = 7, 8, 9. If no cells are specified, **do all 1–9**.

### Orientation (all stages)

| Rule | Detail |
|---|---|
| **Default** | `orientation='portrait'` (9:16) whenever the user does not ask otherwise |
| **Exception** | User **explicitly** asks for landscape / 16:9 / widescreen → `orientation='landscape'` |
| **Forbidden** | Switching to landscape because the reference is landscape, or because the platform default is `landscape` |
| **Pipeline** | Once a job is landscape, i2v and mix_clips keep the same `orientation` |

## Collaboration

### Grid / split: pause at the boundary

At stage end: **same turn**, call only `set_guidance` (no other tools), then wait.

1. Grid done → split? / regenerate? / more fail?
2. Split done → **Generate video**? / re-split? / regenerate grid?

### User chooses Generate video: run stages 3→4 with no extra confirm

Any of these enter the pipeline: generate video, make videos, animate all, ref2v, animate cells, 做视频, 生成视频.

1. **Not split yet** → split 9 cells per `stage-split.md`, **do not** `set_guidance`, go straight to animate
2. **Cells**: user-specified, or **all 1–9** if unspecified
3. Per-cell `i2v` via `stage-video.md` (`duration=3`)
4. **≥2 successful clips** → **immediately** `mix_clips` per `stage-mix.md`. **Do not** ask whether to stitch
5. **Only 1 successful clip** → deliver that clip; no mix
6. After the compilation (or t