Clay World Short · skill pubblica Open Skill
Clay World Short
**Clay World Short · One Figure, One Diorama, One Tactile Cut · ref2v Native · Foley Kept** *The agent skill for creators who want a finished **claymation stop-motion short**—not a photoreal live-action, a LEGO brick film, a still portrait, or a first-frame freeze of a lookbook plate* **Clay World Short** is a production-native Agent Skill for Digen. It orchestrates a complete **reference-to-video (ref2v)** pipeline for one claymation episode: **Idea (or your real photo) → theme lock → clay identity plate + empty miniature → text mini-script → 10s ref2v clips → one mute-free join** No storyboard grids. No contact sheets. No TTS, no `audio_url`, no generated BGM. The lead stays a **handmade clay figure**. Beats are **tactile**. Speech is **off**—clip foley stays on the join. Costume stills are identity plates, not opening keyframes. The film starts **already inside** the clay world. ## What you get | For creators | For the production stack | | --- | --- | | **Beginner-first** — no photo? the agent invents a clay-figure still from the idea you clicked and runs toward the cut | **Opinionated pipeline** — language lock, identity-plate gate, duration=10, ref2v packing, and unmute join encoded in-skill | | **Photo-first** — your person / place / object is the lock; no photo is a generate path, not a dead end | **ref2v-first** — no accidental i2v / t2v / fl2v / director / first-frame chains; no physical model names | | **World that reads on sight** — rounded clay, matte sculpt marks, one miniature, one through-object in the hands | **Two-ref binding** — `input_urls` = clay figure, then empty set; neither still is the first or last frame | | **Dynamic ideas** — figure + diorama + tactile-beat buttons written for this turn, not a canned trio | **Six-section prompts** — `[reference generation]` first, then a compiled English clip body from the mini-script | | **Delivery that continues** — after clip 1, buttons to finish / change a beat / redo; after the cut, next moves in your language | **Progressive disclosure** — craft lives in `references/`, not bloated context | | **Plain creator voice** — short lines about this figure and this miniature; no pipeline jargon unless you ask | **Marketplace-ready** — clear triggers, defaults, and tool contracts | ## Pipeline | Stage | Output | | --- | --- | | **Intake** | Lock language + **16:9** + default **30s** (3×10s). Person photo → bind that face. Place photo → clayify the empty set. No photo + idea chosen → **one `t2i`** clay figure (look + clay in one still). Vague ask → clickable ideas, then generate | | **Identity** | Clay-figure still as a **plain-background plate** (not standing in the diorama). Upload path: `i2i` clay conversion onto the person photo only | | **Set** | Empty clay miniature (`t2i`, or `i2i` from a place photo, `locations`) — no lead in the plate | | **Text chapters** | Logline + **one through-object** + Arrive / Craft / Payoff (default 3×10s) — identity, tactile beat, screenshot hero pose. No grids | | **ref2v** | One **10s** clip per chapter. Tool **must** pass `duration=10`. Prompt is the six-section pack; `summary` starts with `[reference generation]`. Figure + set as **references**, not keyframes | | **Join** | `mix_clips` once, `volume=1`. **No mute. No music tool. No BGM overlay.** 10s-only skips the join | ## Production stack | Parameter | Default | Notes | | --- | --- | --- | | **Episode** | **30 seconds** | 3 clips × 10s. Optional 10s (1 clip) or 20s (2 clips) if you ask | | **Engine** | `ref2v` · `reference_to_video` | Bare default channel. Forbidden: `i2v` / `ai2v` / `t2v` / `fl2v` / `director` / physical names such as `rm3.5-ref` | | **Clip duration** | **10 seconds** | **Always set `duration=10`.** Prompt text “10-second” does not set length. Omit → platform 5s | | **Video orientation** | Landscape · **16:9** | Portrait only if you ask. Figure still, set still, and video share the lock | | **Image refs** | Clay-figure plate always; empty miniature always | Figure = `t2i` `characters` (no person photo) or `i2i` from upload. Set = `t2i` / `i2i` `locations` | | **Ref order** | Figure → empty set | Identity plate, not a first-frame. Forbidden: empty `input_urls`; single-image calls; empty set as opening frame; frozen replica of either still | | **Identity plate** | **On, every no-upload / clay-convert still** | Clear face, clay body, **plain studio background**. Forbidden: figure already inside the diorama in the still | | **Clay lock** | **One handmade world** | Rounded matte clay, stop-motion feel. Forbidden: photoreal skin, pores, hair strands, DSLR look, mixing real materials with clay | | **Lead lock** | **Same identity after lock** | Face shape, hair mass, clothes colors stay. Forbidden: swapping people, aging a child, sexualizing anyone | | **Speaker lock** | **N/A — no speech** | No dialogue, no VO, no `generate_tts`, no burned-in captions | | **Dialogue** | **Off** | Story is visible verbs (pour, press, place, steam), not subtitle jokes | | **Picture audio** | **Clip foley, kept** | Light tactile foley from the video. Join does **not** mute. No `generate_music` | | **Cut grammar** | **3×10s then join** | `mix_clips` `volume=1`. Forbidden: mute join; second pass overlay BGM | | **BGM** | **Off** | No music tool. Do not ask the video model to invent a score | | **Mini-script** | **Logline + through-object** | Arrive / Craft / Payoff, three beats per 10s. Failed board = chapters you could swap | | **Confirm** | **After clip 1** | Default pause so you can confirm the figure. One-shot / generate-all skips that pause | | **Buttons** | **Written this turn** | Not a fixed catalog. Ideas, sets, and next moves named from this figure and this object | | **Language** | **User-typed lock** | Replies, buttons, todos, asset names in your language; image/video prompts stay English | ## Modes | Mode | You bring | Agent delivers | | --- | --- | --- | | **Full auto-run** | “Make a clay world short” / generate all | Invent a figure + miniature if needed → stills → 3 clips → unmute join | | **Idea-first** | One-line premise (“rainy cafe, warm mug at the window”) | One `t2i` clay figure → empty set → mini-script → cut | | **Photo-first** | Your person / place / object photo | Person: bind that face, `i2i` to clay. Place: clayify the empty set. Object: repeating prop | | **Pick-an-idea** | Nothing locked yet | 2–4 concrete figure+diorama+object buttons in your language; click → generate that figure | | **Set swap** | “Make it a night market instead” | New empty miniature; same locked clay figure | | **Single-clip** | “Just 10 seconds” | One ref2v clip with `duration=10`; skip the join | | **Single-stage edit** | “Redo clip 2” / “Change the cafe” / “New object” | That stage only, then remake the affected clip / cut | | **One-shot** | “Finish in one go / no confirm” | No pause after clip 1; still pause on the finished cut | | **Lengthen** | “20 seconds” / “30 seconds” | 2 or 3 × 10s chapters, then one unmute join | ## Example prompts **Full auto-run** > Make a clay world short. Landscape, 30 seconds. **Idea-first** > Rainy cafe. A clay figure wants one warm mug at the window before the rain thickens. 16:9. **Photo-first** > This is me. Turn it into a claymation gardener watering a terrace. Landscape. **Pick-an-idea** > I don’t have a photo. Give me a few clay-world ideas and just start. **Set swap** > Same figure. Change the set to a night-market stall. **One-shot** > Generate everything in one go. Don’t stop after the first clip. **Single-stage edit** > Redo the middle clip — the beat should be the pour overflowing, not looking around. ## When to use - Landscape claymation / stop-motion / miniature-diorama shorts: one figure, one set, one object in the hands - You uploaded a real photo of a person, place, or object, **or** you only typed a scene - You want to change the figure, rewrite chapters, redo one clip, or remake the cut - Creators who want a finished 30-second join with original clip sound, not a mute beauty reel > **Not the right tool when** you need photoreal live-action, LEGO brick films, 2D anime, a product ad that must stay photoreal, a spoken / lip-sync track, a generated song overlay, storyboard grids for client sign-off, or real-world harm. ## Invoke ``` /clay-world-short ``` Or: *"Turn this into a 30-second landscape claymation — same face if I uploaded one, one miniature, one object in the hands, ref2v, duration 10 per clip, keep the clip sound, no mute mix."* **Your photo · Your miniature · One coherent clay world.**
ShortDrama
21 4 7
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3325120/2026-09-18/bjsewgx8.png
# Clay World Short Turn a real photo or a typed scene into a **landscape claymation short**. One clay figure lives in one miniature diorama. Surfaces look hand-sculpted: rounded edges, matte clay texture, warm soft light, stop-motion pacing. End on a screenshot-worthy clay pose. This Skill has **no** structured project state. Keep facts in the reply or `create_document`. Reuse canvas URLs. **Video channel (hard):** every clip uses `generate_video(model='ref2v', gen_mode='reference_to_video', duration=10, orientation='landscape')` — the **bare `ref2v` default channel**. **`duration=10` is a required tool argument.** Writing "10-second" only in the prompt does **not** set length. If `duration` is omitted, the platform default is **5 seconds**. Forbidden: omitting `duration`; passing `duration=5`; `i2v` / `t2v` / `ai2v` / `fl2v` / `director` / physical model names such as `rm3.5-ref`. `gen_mode='reference_to_video'` must be passed with `ref2v`. `input_urls` = **clay-figure still, then empty clay set** (at least 1 image; 2 required when both exist). Never put an audio URL in `input_urls`. This is **reference-to-video**. The prompt is the **six-section pack** in `references/04-video.md`; `summary` **must start with** `[reference generation]`. Forbidden: a 5-sentence mood blurb; treating any input still as the first/last frame; `[image to video]`; first-frame / keyframe completion. **Zero CJK in this package.** Skill files stay English (handbook only). User-facing runtime copy follows `user_lang`. Image and video `prompt` fields stay English. Image prompts follow packing in `references/02-assets.md` (short English). Video prompts follow the six-section pack in `references/04-video.md` (English; compile from the mini-script, not a mood paragraph). ## Global Language Lock (mandatory, highest priority) Follow the **user's typed sentences**, not platform chrome. **Chrome trap:** strip the slash-invoke line, `<untrusted_skill_content>`, `[session_assets]`, tool traces, and injected English such as "The synthesis failed…". **CJK wrapper ≠ `zh-CN`. English Skill dump ≠ `en`.** Count Han only in leftover `typed_user`. Zero Han → not Chinese (English words → `en`) and **zero CJK** in every user-visible string. Majority Han → `zh-CN` / `zh-TW`. Translate user copy at runtime — never paste handbook English, and **never put Han characters into this Skill**. **Two layers:** `user_lang` (everything the user can read) and MODEL English (`generate_image` / `generate_video` `prompt`). Algorithm: `references/00-language.md`. On the **first user message**, silently `read_skill_file` **only** `references/00-language.md` once, then lock `user_lang` **before any user-visible sentence and before any `generate_*` / `create_document`**. Do not re-detect after tools, English files, `[session_assets]`, or "continue". Do **not** also load later stage files on that turn. The first user-visible sentence must already be in `user_lang` and about **this photo / this clay figure / this diorama**. | Layer | Scope | Language | |---|---|---| | **User-facing** | Replies, buttons, todos, asset `name`, canvas body | `user_lang` | | **Image / video prompt** | `generate_image` / `generate_video` `prompt` | **English always** | `zh-CN` vs `zh-TW` is a lock. One turn = one `user_lang`. If a mix is found, rewrite before send. Platform English never re-detects. Switch `user_lang` only when the user explicitly asks to change reply language. ## When to use - The user wants to enter a clay world, make a claymation / stop-motion / miniature diorama short - The user uploaded a real photo of a person, place, or object, **or** only typed a scene with no photo - The user only wants to change the figure, redo one clip, or remake the cut Do not use for: photoreal live-action, LEGO brick films, 2D anime, product ads that must stay photoreal, or real-world harm. ## Intent lock (first, every run) Before tools, write one internal sentence: **what this user wants this turn**