Pov Maker · Open Skill 公開スキル

Pov Maker

**Pov Maker · Sight to Still to Screen · ref2v Native · Mix-Free** *The agent skill for creators who want a finished **first-person cut**—not a follow-cam of a character, a mute beauty reel of someone else, or dangling FPS arms that fight the video model* **Pov Maker** is a production-native Agent Skill for Digen. It orchestrates a complete **reference-to-video (ref2v)** pipeline for one first-person episode: **Idea → I-see story → text shots → scene still (+ POV still only if the body is in frame) → one 10s ref2v final cut** No storyboard grids. No contact sheets. No TTS, no `audio_url`, no six-clip stitch. The camera **is your eyes**. Speech is **off**—ambience, foley, and instrumental score are generated **inside the video** from the six-section prompt. Own body appears only as a closed `body_reveal` lookup, or not at all. ## What you get | For creators | For the production stack | | --- | --- | | **Beginner-first** — no idea? the agent invents a 10-second I-am-seeing scene and runs to the cut without stopping | **Opinionated pipeline** — FPV LOCK, closed body-reveal table, still gate, and six-section prompt encoded in-skill | | **Sight-first** — synopsis and shot list as **what occupies the frame through my eyes**, not a person being filmed | **ref2v-first** — no accidental i2v / t2v / fl2v / storyboard / per-shot video chains | | **Identity of place, not a portrait** — empty-world scene still in every render; optional first-person still only when the lookup shows self-parts | **Multi-ref binding** — `input_urls` = scene, then POV still if it exists; never a lookbook as the POV slot | | **Living FPV grammar** — world filling the frame, approach without arms, hands only already working a mid-frame object | **Dense prompts** — FPV LOCK first + `body_reveal={tokens}` + six sections + 3–6 `[Shot N]` optical windows | | **Delivery that continues** — after the cut, buttons are **Edit shots / Regen scene / Regen video / Next episode**, already in your language | **Progressive disclosure** — craft lives in `references/`, not bloated context | | **Plain creator voice** — short lines about what I see and what happens; no pipeline jargon unless you ask | **Marketplace-ready** — clear triggers, defaults, and tool contracts | ## Pipeline | Stage | Output | | --- | --- | | **Intake** | Lock language + **9:16** + **10s** + 1 episode (no orientation question) | | **Story / world** | First-person synopsis as **sight**; host analysis; closed `body_reveal` tokens | | **Text shots** | 3–6 optical cards — world first, then only legal self-parts; durations sum to 10 | | **Visual assets** | Scene still (`t2i`, empty space from camera height). POV still (`i2i`) **only if** Shot 1 `body_reveal` is not `none` | | **ref2v** | One six-section prompt → **10s** `final_videos` cut. Diegetic + score generated in-clip | | **Sequel** | Next episode from end-state + seed — reuse host / room when unchanged, new sight-arc, one new 10s video | ## Production stack | Parameter | Default | Notes | | --- | --- | --- | | **Episode** | **10 seconds** | One `ref2v` video. Overflow → next-episode seed, not jammed in. User asks longer → explicit duration or more 10s chapters | | **Engine** | `ref2v` · `reference_to_video` | Multi-ref identity lock. Forbidden: `i2v` / `ai2v` / `t2v` / `fl2v` / `director` | | **Clip duration** | **10 seconds** | Always set. Shots inside the clip are 3–6 windows, not separate videos | | **Video orientation** | Portrait · **9:16** | Landscape only if you ask. Scene still, POV still, and video share the lock | | **Image refs** | Scene always; POV still only when the body is legally in frame | Scene = `t2i` `locations`. POV still = `i2i` from the scene, **not** `characters` (that biases portraits) | | **Ref order** | Scene → POV still (if any) → optional uploads | Never a full-body portrait as the POV slot. Clothing uploads bind material only | | **FPV LOCK** | **On, every prompt** | Camera = my eyes. Forbidden: follow-cam, OTS, “POV of a person”, zombie walk, FPS HUD hands | | **Body reveal** | Closed lookup | `gaze_pitch × posture` → tokens only. Walking + looking ahead = `none`. Hands only already on the working object | | **Speaker lock** | **N/A — no speech** | No on-camera talker, no off-screen VO, never both. The host is behind the lens | | **Shot grammar** | **Optical first person** | World occupies the frame. Cut only inside one 10s clip. No quota cuts to “make it POV” with dangling arms | | **Dialogue** | **Off** | Forbidden: `generate_tts`, `audio_url`, `<d>`, spoken lines, burned-in captions | | **Picture audio** | **Diegetic + instrumental** | Ambience, foley, and score described in the six-section prompt and generated in-clip | | **Cut grammar** | **Inside one clip** | `[Shot N]` blocks in one 10s render. No storyboard grids, no per-shot i2v chain, no `mix_clips` | | **BGM** | **In-prompt, not mixed** | Instrumental in `non_diegetic_music`. No music tool. No 15% mix pass | | **Confirm** | **After delivery** | First episode auto-runs; pause only if you ask, then buttons after the video | | **Series** | **Next episode** | Reuse host / look / room when unchanged; new scene still if the place changed | | **Language** | **User-typed lock** | Replies in your language; image/video prompts stay English; one script per visible string | ## Modes | Mode | You bring | Agent delivers | | --- | --- | --- | | **Full auto-run** | Nothing — or “make a POV” | Invent an I-see scene → story → shots → stills → one 10s cut | | **Idea-first** | One-line premise or mood | Expand to a 10s sight-arc → full pipeline | | **Script-first** | Plot, place, or “what I see” draft | Lock host + `body_reveal` → shot cards → stills → ref2v | | **Reference-driven** | Room / clothing photos + brief | Bind uploads for place and material; do not copy a standing pose | | **Eyes-only** | “No hands / no body in frame” | `body_reveal=none`, scene still only, world fills the frame | | **Embodied interact** | “I push the door / part the vines” | Hands already on the object in the working plane — not a zombie reach while walking | | **Single-stage edit** | “Redo the shots” / “Change the place” / “Regen video” | That stage only, then remake the cut | | **Same-thread sequel** | “What happens next?” / Next episode | Continue from end-state · URI reuse · one new 10s video | | **Lengthen** | “This episode should not be only 10 seconds” | Prefer more 10s chapters; single longer clip only if you name the duration | ## Example prompts **Full auto-run** > Make a first-person POV short. Portrait, 10 seconds. **Idea-first** > Rain alley at night. I see the metal door ahead, neon in the puddles. Portrait, first person. **Eyes-only** > Dusk rooftop, drone view, no body in frame. First person, 9:16. **Embodied interact** > Jungle temple. My hands are already parting the vine curtain. No zombie walk, no arms reaching forward. Portrait POV. **Reference-driven** > This photo is the jacket material. This photo is the room. Make a 9:16 first-person short. The camera is my eyes. **Script-first** > Shoot exactly this “I see…” copy. Do not invent a person for the camera to follow. 9:16, episode 1. **Series** > Continue the last story — what is behind the door after it opens. ## When to use - Vertical first-person shorts: I see the space, I approach, I work a reachable object - One-line ideas, pasted “what I see” copy, or room / clothing stills - Hosts that are not always human — cabin, drone, animal, ghost, object — with anatomy-honest limbs - Creators who want a finished 10-second cut, then the next chapter in-thread - Anyone tired of “POV of a woman walking,” storyboard grids, or zombie-arm FPS cameras > **Not the right tool when** you need a 60-second six-clip mix, a spoken product ad with lip-sync, lip-sync to a generated song, storyboard grids for client sign-off, a full-body character lookbook, external TTS as the primary track, horror / gore, or a third-person drama the camera follows. ## Invoke ``` /pov-story ``` Or: *"Turn this into a 10-second portrait POV — the camera is my eyes, place locked, one ref2v cut, no follow-cam."* **Your eyes · Your space · One coherent ten seconds.**

@jimi

ShortDrama

163 14 99

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3325120/2026-08-18/9tw7rau3.jpg

---
name: pov-story
display_name: Pov Maker
description: 'Use when the user wants Pov Maker / a first-person POV story short video (from idea to final cut: story, shots, scene still, first-person still, video), including my viewpoint; also use when they only want to advance or change one stage (revise shots, regenerate video, make a sequel)'
allowed-tools:
- generate_image
- generate_video
---

# Pov Maker

Turn a one-line idea into a **first-person POV** short video. There is no structured project state; progress lives in the conversation, canvas assets, and `write_todos`.

Default: **one 10-second video per episode**, with **3–6 shots** inside it. Shot craft goes into the multi-parameter `ref2v` prompt as `[Shot N]` blocks. **Do not** generate storyboard grids. **Do not** generate a full-body character portrait. **Do not** use TTS, music tools, or `mix_clips`. The `ref2v` video is the final cut; ambience, foley, and score are written into the video prompt.

## FPV LOCK (highest visual rule; every stage)

First-person is **sight**, not a genre tag and not a shot of a person.

**What it is:** the finished frame is the world as seen from inside the host. The host is behind the lens. If a human could not see it without turning into a second camera, it is not first-person.

**What it is not:** a movie about a character; a tracking shot; over-the-shoulder; the back of a head; a standing portrait; "POV of" someone.

Paste this English block as the **first paragraph** of every `generate_image` prompt and every `generate_video` six-section body. Do not paraphrase it away:

```text
FPV LOCK: This picture is the view out of my own eyes. I am the camera, not a subject on screen. The world occupies the frame. There is no actor who is me standing, walking, or facing the lens. Forbidden: third-person, follow-cam, over-the-shoulder, back of a head, full-body person as me, selfie, POV of a character, zombie walk, shambling, stiff dangling arms in the lower third, both arms reaching forward while walking, FPS HUD hands parked at the bottom. Own body appears ONLY as body_reveal tokens, or not at all, attached to me behind the lens. When walking and looking ahead, no own arms in frame. Hands enter only already working on a mid-frame object. Locomotion means the world streams toward a living human gaze, horizon mostly stable, not a filmed walker and not a corpse shuffle.
```

Then state `body_reveal={tokens}` (from the Stage 2 table) so the model knows what of me, if anything, is in frame.

**Anti-zombie (same rank as FPV LOCK):** empty or reaching arms in the lower third of a 9:16 walk **is** the zombie camera. Do not pick `down_mild` on a walk just to unlock `forearms+hands`. Eye-level locomotion stays `body_reveal=none`. Hands appear only when `pose_task` is already in contact with a reachable object on the working plane — never "reaching both hands forward" and never "hands at lower rails / lower frame edges" as a rest pose.

**Reject and rewrite** any synopsis, `first_frame`, image prompt, or `[Shot]` that could be filmed by a second camera pointed at a person, **or** that parks idle arms at the bottom of the frame during a walk. Rewrite until the grammatical focus is the world ahead (or own-body tokens from the table doing a real job).

User-facing story and shot list: first person as **I see / the space ahead**. Forbidden as the visual subject: a man, a woman, a girl, the protagonist, the character, he/she walking.

## Global Language Lock (highest priority; branches must not weaken it)

Two language layers apply at once and must not overwrite each other. This Skill document and every file under `references/` are **English only**. Never write Han / CJK characters into this Skill package. **Forbidden:** treating the document language, tool English, or phrase-table English as the user-facing language — or writing model prompts in the user's language.

### 0. First-turn lock (must finish before the first reply)

- Lock `user_lang` from the *