@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/framewright/references/world.md

@@ -1,215 +0,0 @@
-# One continuous world
-
-Read before building a film without cuts. Everything here is implemented in `assets/world.html`;
-`init.sh --world` scaffolds it as `index.html` with a two-plate demo that renders out of the box.
-`examples/honeybee` in the repository is a finished 56-second film built this way.
-
-## 1. When
-
-A plate film cuts between pictures. A world film is one take: the camera never cuts, a line
-never lifts, the viewer travels. Choose it when the logline is a journey or a growth: a line
-that draws a life, a flight from a detail out to a landscape and back, a map that unrolls, a
-machine assembled part by part. Keep plates for bulletins, parodies, lists and anything cut on
-the beat.
-
-The price: scenes stop being independent. A change early in the film moves everything after it
-(section 9), and review happens by sections, not by plates.
-
-## 2. Anatomy
-
-Plates keep their lengths on the beat grid, but each one gets a builder, and all of them share
-one draw function:
-
-```js
-plate('spark', {len: 2*BAR, build(W, {pen, b, R}) {
-  pen.hold(b(1));                                  // b(x): the frame of beat x of this plate
-  pen.along(ring, b(4), 'sio', {keep: 0.85});      // draw the ring, arrive on beat 4
-  camKey(W, b(0), 0, 0, 34); camKey(W, b(8), 0, 0, 50);
-  title(W, b(2), b(7.2), 'SMALL LINE', 'Big line');
-}});                                               // the draw function defaults to drawWorld
-```
-
-`buildWorld(seed)` runs once per seed and is cached. It computes `W.start[name]` from the plate
-lengths, runs the builders in order with one pen that carries over from plate to plate, sorts the
-camera keys, cuts the line into drawing chunks and checks the titles. `drawWorld(S)` renders any
-frame from `W` alone, so a frame stays a pure function of (frame, seed, width).
-
-| builders append | how | drawn |
-|---|---|---|
-| the line | `pen` moves (section 3); a second line through `extraPen` | light layer |
-| camera keys | `camKey` (section 4) | every frame |
-| titles | `title` (section 7) | over everything |
-| light events | `W.glows.push({f0, f1, draw(lg, f, k)})`: flashes, sparks, a glowing drop | light layer |
-| decor | `W.sets.push({f0, f1, draw(g, f, k, cam)})`, may outlive its plate | under the light |
-| hidden or fading parts of the line | `cover`, `dim` (section 5) | light layer |
-| values over the whole film | functions on `W`: `W.spot(f)`, `W.headGlow(f)` | where used |
-
-Time inside a builder is in beats: `b(1.5)` is one and a half beats into the plate, `b(8)` two
-bars. Another plate's time is `W.at('name', beats)`. A moment the sound must hit goes into
-`W.cues.push({name, f: b(x)})`, next to the glow or stamp it sounds with. Never write global frame
-numbers. Randomness comes from `R`, seeded by the plate name. `look.mjs info` prints the plates
-with their start frames and the size of the world.
-
-`R` belongs to the builder: take every random value while building (a tilt, a jitter, a choice)
-and keep it in a variable the draw function reads. A draw function that calls `R` itself advances
-the generator on every render, so the same frame comes out different each time; `check.mjs`
-reports such a frame as changing in every render order. This bit a real film through the tilt of a
-few sticky notes.
-
-## 3. The line
-
-The line is five arrays: `X, Y` in world units, `T` the frame at which the head passes the point,
-`A` the memory floor (how bright a stroke stays after the head has passed, 0..1), `Wd` a width
-multiplier. The pen writes them:
-
-| call | does |
-|---|---|
-| `pen.along(pts, t1, ease, o)` | follow a polyline from the current point, arrive at frame `t1` |
-| `pen.to(x, y, t1, ease, o)` | a straight stroke |
-| `pen.bez(c1, c2, p, t1, ease, o)` | a cubic Bézier from the current point |
-| `pen.hold(t1)` | stay still until `t1` |
-| `pen.lift(x, y, t1)` | move to `(x, y)` by `t1` without drawing: tally marks, crosses, checks, letters |
-
-Options: `keep` (the memory floor, a number or a function of the stroke fraction `u`), `width`
-(the same), `step` (resampling step in world units: 0.5, or 1–2 for long flights seen from far),
-`warp(u, x, y)` returning an offset (a tremble that grows toward the end, a wiggle). Eases: `sio`
-(default), `io`, `in`, `out`, `sin`, `sout`, `lin`.
-
-Points are resampled at a constant step and each point's time comes from the inverse of the ease,
-so the head moves along the whole polyline with one ease however its vertices are spaced. `hold`
-writes a point, so the head does not creep into the next stroke. A stroke that arrives before the
-pen's current time means the previous stroke overran; the pen warns in the console and draws it
-at once. `buildWorld` also warns when a plate's strokes run past the plate's end. A lifted pen
-writes points of width 0; the line is cut so that no drawing chunk mixes ink and lifted pen, and
-lifted chunks are never stroked.
-
-Drawing: the head at frame `f` is found by binary search in `T` and interpolated between two
-points. The finished part is cached as `Path2D` chunks of 20 points; only the chunk that holds the
-head is rebuilt per frame. Every chunk is stroked twice, a wide faint glow (9 px, alpha × 0.2) and
-a bright core (2.3 px), with alpha `A + (1 − A)·exp(−age/τ)`, τ ≈ 75 frames: a fresh stroke
-burns bright and settles to its floor. Floors: 0.15–0.35 for travel, 0.5–0.85 for strokes that
-must stay readable. A 56-second film had 23 000 points in 1 150 chunks and 48 camera keys; a
-1920 px frame with its PNG took 0.3 s.
-
-## 4. The camera
-
-`camKey(W, f, x, y, z, {w, lag, ox, oy, e})`: at frame `f` look at `(x, y)` with `z` world units
-across the short side of the frame. Because `z` counts the short side, the same key frames the
-same subject in 16:9 and 9:16.
-
-- Between keys `x, y` follow the ease of the key you move toward, and `z` is interpolated in log
-  space, so a zoom has an even pace. A linear `z` rushes at the close end and crawls at the far end.
-- `w` (0..1) pulls the aim toward the head as it was `lag` frames ago, shifted by `ox, oy`.
-  0.3–0.85 with a lag of 7–10 keeps a travelling head in frame and a little ahead of the centre.
-  `w` blends between keys like everything else, so following fades in and out.
-- The result is smoothed with a Gaussian over ±18 frames (σ = 6; `z` averaged in log space).
-  Corners between keys become curves, starts and stops ease, and the camera settles on a key up to
-  12 frames after it: put the key earlier when an event must be framed on its frame.
-- To track a falling or thrown object, push a key every 4 frames along its trajectory with
-  `e: 'lin'`; the smoothing does the rest.
-- `k = SHORT / z` is logical px per world unit. Whatever must keep its size on screen through a
-  zoom (line widths, dots, hairlines of decor, glow radii) is given in px and divided by `k`.
-- Pace from a real film: a zoom of about ×2 per second reads as the world opening, ×5 per second
-  and faster as a dive. A hold drifts at most 20 % of the frame per second.
-
-Before the first key and after the last the camera stands still; with no keys it frames `z = SHORT`
-around the origin.
-
-## 5. Light, or ink on paper
-
-The line, the head and every glow are drawn in world coordinates on an offscreen canvas of the
-frame's size, then added onto the picture twice with `lighter`: through `filter = blur(W/170 px)`
-at alpha 0.85 (bloom), then sharp. Light adds to a dark picture and never muddies it; keep the
-ground dark wherever light must read.
-
-On a light ground (paper, a lightbox, a whiteboard) added light disappears: a red line added to
-white stays white. Set `LINE_ON = 'paper'`: the line is stroked onto the picture itself with normal
-compositing, after the decor, and the wide faint pass becomes a soft bleed of the ink. Covers work
-only on the light layer. Give the tool a body with `W.drawHead = (g, x, y, S, f) => ...`, called at
-the head in screen coordinates after everything but the titles: a grease pencil tilted like a hand
-holds it, a marker, a needle. It works in both modes.
-
-- The head: a soft dot of 20 px with a 5 px core, scaled by `W.headGlow(f)` (a breath before the
-  first stroke, dimmer while a character sits on the head).
-- Occlusion: `cover(W, f0, f1, cut)` hides everything the light layer holds from before the call.
-  That older part of the line is drawn, `cut(lg, f, k)` erases it with `destination-out` (fill the
-  shape of the lid, the wall, the fog, with the cover's opacity), then newer strokes are drawn on
-  top. The visible lid itself is decor in `W.sets`. Glows drawn before the line are erased too.
-- Dimming: `dim(W, i0, i1, f0, f1, to)` fades points `i0..i1` of the line to `to` of their
-  brightness between frames `f0` and `f1`: a memory that fades, a broken part, a past that recedes.
-  Take the indices from `W.path.X.length - 1` in the builder where that part begins and ends.
-- Both add chunk breaks at their indices, so a hidden or dimmed range starts and ends exactly
-  there. Use them rather than raw index ranges.
-
-## 6. Decor and the spotlight
-
-Decor is drawn in world coordinates under the light and culled with `viewRect(cam, k, margin)`.
-Dots and hairlines get screen sizes (`px / k`), big objects world sizes.
-
-In a close-up the set around the subject competes with it. Dim the decor with a spotlight measured
-on screen: multiply its alpha by `spotlight(W, f, cam, k, x, y)`, which is 1 inside 30 % of the
-radius and 0 beyond `W.spot(f) · LH`. About 0.6 in an intimate scene, 3 or more when the world
-opens; animate `W.spot` with the story. Measured on screen, the pool keeps its size through a zoom.
-
-Decor that must read at every zoom: grids of jittered dots at several scales (spacing ×4 per
-scale), each shown only while its dots stand at least ~40 px apart on screen. The demo does this.
-A film that travels through distinct places (a room, a house, a field) gives each place its own
-decor at its own scale instead.
-
-## 7. Titles
-
-`title(W, f0, f1, small, big, o)`: a small letter-spaced line over a big one, in screen space, on
-the baseline `TITLE_Y`: 80 % of the height in 16:9, 70 % in 9:16, because the bottom fifth of a
-vertical video belongs to the platform's captions and buttons. `small` and `big` may be functions
-of the frame (a counter). A title fades in over 14 frames, out over 16, rises 12 units, and shrinks
-to fit 84 % of the width. Two titles in the same place at the same time read as one smudge: end
-one before the next begins. `buildWorld` warns about overlaps, and `look.mjs` prints the warning.
-
-Over a light or busy picture give the title a plate: `{plate: 0.72}` draws a dark rounded plate
-behind the block. `TITLE_SCRIM` (0..1) darkens the lower half of the frame while a lower title is
-up, so the next scene waiting below does not crowd the text.
-
-A cover on frame 0, for feeds that show the first frame as the preview: `FADE_IN = 0`, a title with
-`f0` below zero (fully visible on the first frame), and strokes that already exist when the film
-starts. Put `PEN_START` at the first stroke with a negative time (`[x, y, -60]`) and draw that stroke
-to a time below zero: a circled photo is on the cover, and the line leaves it once the film runs.
-Then shoot frame 0 at full size: it is also the poster.
-
-## 8. Review
-
-- Sheets per section, not only for the whole film:
-  `FROM=480 TO=720 node scripts/look.mjs sheet 12 480 7 shots/sheet-part.png`. The whole sheet
-  shows the story; a section sheet shows whether the camera stalls or jumps.
-- At every plate boundary shoot the last frame of one plate and the first of the next: the line and
-  the camera continue, nothing jumps.
-- An empty scene is a scale problem. Bring the camera closer (smaller `z`) or make the object
-  bigger; more detail makes a scene busier, not fuller.
-- A busy close-up: lower `W.spot`, darken the set, before removing anything.
-- `node scripts/check.mjs` before the full render: the world is cached per seed, and any state a
-  builder or a draw function keeps outside `W` shows up there as a frame that changes with render
-  order.
-- The vertical cut loses the sides of the landscape frame and gains top and bottom. Render its own
-  sheet with `AR=9:16`; if the subject leaves the frame, add `oy` or `w` on the keys of that
-  section rather than a second camera.
-
-## 9. Changing a world film
-
-- Retiming a plate moves every later plate in time, not in space: builders count in `b()`, so
-  their strokes keep their shapes.
-- The pen carries over. A plate that starts with a relative move (`pen.x + 40`) starts wherever the
-  previous one ended; one that goes to absolute world points does not care. Anchor each plate's
-  key moments to world points.
-- Inserting a plate changes the first stroke of the next one: reshoot its first frames.
-- Randomness stays per plate name; world-wide decor is seeded on its own.
-- After any change: the section sheet, the boundary frames, then curves and sound regenerated
-  (`make.sh` exports the curves before `audio.mjs`).
-
-## 10. Sound from the picture
-
-The world knows where the head is and how fast it moves on screen, so the sound can follow the
-picture instead of a hand-written cue list. `RISO.curves()` returns plate starts and per-frame
-values (`sp` head speed in px per frame, `pan` head position −1..1, `z`, `w` the width at the
-head, 0 while the pen is lifted), plus `strokes`, the frames where the pen touches down after a
-lift, and `cues`, the moments builders marked. `scripts/export-curves.mjs` writes them to
-`curves.json`; `audio.mjs` takes its timeline from there, drives a voice with `follow()`, puts a
-tap on every touch-down and a hit on every cue. Recipes: `references/audio.md`, section 7.

package/skills/gbro-collage-broll/SKILL.md

@@ -1,435 +0,0 @@
----
-name: gbro-collage-broll
-description: "将约 5 秒口播文稿、观点句或抽象概念做成高级 editorial halftone paper-collage / 半调纸拼贴 B-roll。用户说“collage b-roll”“纸拼贴 b-roll”“半调拼贴”“拼贴风格配画面”“用这段文稿做拼贴动画”“gbro-collage-broll”，或希望把一句文稿转成拼贴视觉隐喻时，必须使用此 skill。强制采用三阶段审批：先只提视觉隐喻，用户确认后才生成彩色拼贴静帧，静帧再次确认后才默认调用 Gemini Omni Flash 生成首尾帧组装动画。默认视频模型固定为 gemini-omni-flash-preview，不再默认使用 Veo；只有用户明确指定其他模型时才切换。 Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "pyang5166/gbro-collage-broll"
----
-
-# gbro Collage B-roll
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-把一句约 5 秒的口播压成一个 sharp visual idea，再做成高级编辑风纸拼贴组装动画。
-
-默认链路：
-
-1. 只设计视觉隐喻，等待用户确认
-2. 只生成最终静帧，等待用户确认
-3. 自动调用 Gemini Omni Flash 生成视频并完成 QA
-
-这两个确认闸门是工作流的一部分。它们让用户把注意力放在审美和方向上，同时避免错误隐喻或错误静帧直接消耗视频生成成本。
-
-## 首次使用：环境自检
-
-每次触发本 skill 时，进入 Gate 1 之前先运行自检脚本：
-
-```bash
-bash <本skill目录>/scripts/check_setup.sh
-```
-```
-全部通过则直接开始 Gate 1，不要向用户重复配置信息。任何一项失败时，视为首次使用：不进入 Gate 1，先向用户输出下面的配置指南（只列出缺失项），等用户确认配置完成后重新自检。
-```
-
-### 配置指南（按缺失项输出）
-
-1. **GEMINI_API_KEY 未设置**
-   到 [Google AI Studio](https://aistudio.google.com/apikey) 创建 API key，然后写入 shell 配置：
-
-   ```bash
-echo 'export GEMINI_API_KEY="你的key"' >> ~/.zshrc && source ~/.zshrc
-```
-```
-注意：视频生成按量计费，由这个 key 对应的 Google 账号承担。
-
-2. **ffmpeg / ffprobe 缺失**
-   macOS：`brew install ffmpeg`；Debian/Ubuntu：`sudo apt install ffmpeg`。
-
-3. **Python 环境缺失或版本过旧**（需要 >= 3.10）
-   macOS：`brew install python3`；或从 python.org 安装。
-
-4. **共享 venv 未创建**（google-genai >= 2.10.0）
-   征得用户同意后自动创建，无需用户手动操作：
-```
-```bash
-~/hyperframes-projects/.omni-venv/bin/python -m pip install --upgrade "google-genai>=2.10.0" "httpx[socks]"
-```
-```
-5. **不在 Codex 环境 / 没有内置 image_gen**
-   Gate 2 的静帧生成依赖 Codex 内置 `image_gen` 工具。在其他 agent 环境下运行时，向用户说明需要自行提供等效的图片生成方式（或手动提供静帧图片后从 Gate 2 确认继续）。
-```
-
-## 强制审批协议
-
-### Gate 1：隐喻确认
-
-收到文稿后，先提视觉隐喻，不生成图片、不生成视频、不调用任何视频模型。
-
-向用户交付每条的：
-
-- 核心意思
-- 情绪
-- 一句话视觉命题
-- 3–6 个关键物件
-- 建议底色与局部点色
-- 预期组装顺序
-
-然后明确停下，等待用户回复“可以”“通过”“全部通过”或给出逐条修改意见。
-
-如果用户只确认部分编号，只让通过的条目进入 Gate 2；未通过条目继续修改隐喻。
-
-### Gate 2：静帧确认
-
-隐喻确认后，才写 visual spec 和 imagegen prompt，并用 Codex `imagegen` 生成最终静帧。
-
-把原图保存到项目目录，生成带编号的静帧 contact sheet，向用户展示并再次停下。此阶段仍然不调用 Omni Flash，也不生成视频。
-
-如果用户只确认部分静帧，只让通过的条目进入 Gate 3；需要修改的静帧先重生并重新确认。
-
-### Gate 3：视频生成
-
-静帧确认后，不再询问使用哪个视频模型，直接使用本 skill 自带的 `scripts/generate_video.py`，默认调用：
-
-```text
-gemini-omni-flash-preview
-```
-只有用户明确指定其他视频模型时，才覆盖这个默认值。不要自动调用 Veo，也不要把模型选择再抛给用户。
-```
-
-## 成功标准
-
-- 一句话只表达一个清晰隐喻
-- 同一批画面有统一设计语言，但不强制全部蓝底
-- 背景是强烈、平坦、均匀的色场，可按语意变化
-- 主体以黑白 halftone photographic cut-outs 为骨架
-- 关键卡片、按钮、胶片、规则册等允许使用红、黄、青、橙、紫、奶油白等彩色纸张
-- 所有纸片有清晰裁切边、奶油白 keyline、低透明度柔和阴影和纸张颗粒
-- 动作是 assemble-from-empty，而不是轻微漂移、晃动或慢 zoom
-- 无字幕、无口播全文、无 logo、无水印、无 UI
-- 默认交付 9:16、5 秒、720×1280、24fps、无声 MP4
-
-## 什么时候不要用
-
-- 需要精确控制图层、遮挡、镜头穿越或可编辑时间线：改用分层动画工具（如 HyperFrames 类 HTML 渲染视频方案）
-- 只需要视频提示词，不需要生成成片：直接写 prompt 即可，不用走本流程
-- 需要真实人物产品广告或口播演员：不要走本拼贴流程
-- 用户明确要可逐层修改的透明素材：本 skill 默认不拆透明图层
-
-## 默认项目目录
-
-使用北京时间 `Asia/Shanghai` 命名：
-
-```text
-~/hyperframes-projects/YYYY-MM-DD-collage-broll-标题/
-```
-批量项目推荐结构：
-```
-```text
-<project>/
-├── brief.md
-├── visual-spec.json
-├── imagegen-prompts.md
-├── omni-jobs.json
-├── gate2-qa.md
-├── gate3-qa.md
-├── still-contact-sheet.jpg
-├── omni-contact-sheet-all.jpg
-├── video-first-frame-all.jpg
-├── end-frame-comparison-all.jpg
-├── 01-概念名/
-│   ├── omni-prompt.txt
-│   ├── frames/
-│   │   ├── last-frame-original.png
-│   │   ├── first-frame.png
-│   │   └── last-frame.png
-│   └── omni/run-v01/
-│       ├── final-5s.mp4
-│       ├── final-5s-noaudio.mp4
-│       ├── contact-sheet.jpg
-│       ├── video-last-frame.jpg
-│       └── end-frame-comparison.jpg
-└── 02-概念名/...
-
-## Phase 1：设计视觉隐喻
-
-先把文稿压成一个视觉命题。
-
-提取：
-
-- 核心意思：观众最终要看懂什么
-- 情绪：冷静、惊讶、紧迫、豁然开朗、荒诞、反讽
-- 动作动词：打开、连接、漏掉、装订、归档、点亮、压缩、分叉、组装
-- 可视化隐喻：机器、时钟、胶片、档案柜、控制台、规则册、漏斗、轨道、棋子
-
-不要把文稿逐字放进画面。默认一条文稿只做一个隐喻，控制在 3–6 个关键物件；元素过多会让语意变弱，也会让 Omni 组装不稳定。
-
-批量隐喻优先形成前后叙事：例如先表现手工消耗与经验流失，再表现规范沉淀与人机分工。
-
-### Gate 1 输出示例
-
-```text
-1. 核心意思：经验每次都在重复消耗
-   视觉隐喻：熟练剪辑师围着巨大的胶片时钟逐帧裁切，时钟走完一圈却只得到一小段成片
-   关键物件：胶片时钟、剪辑师、剪刀、短胶片
-   色彩：焦橙底，奶油白与浅青点色
-   组装顺序：时钟 → 人物与剪刀 → 胶片 → 最终短输出
-```
-输出后停下等待确认。
-```
-
-## Phase 2：生成彩色拼贴静帧
-
-隐喻确认后，先写自包含的 `visual-spec.json`，再写 imagegen prompt。
-
-### Visual spec
-
-```json
-{
-  "script_meaning": "",
-  "visual_metaphor": "",
-  "style_signature": "flat bold color field, mixed black-and-white halftone cut-outs and colored cardstock accents, crisp cut edges, cream keylines, soft paper shadows, editorial paper collage",
-  "aspect_ratio": "9:16",
-  "color_field": {
-    "background_hex": "",
-    "accent_colors": [],
-    "paper_grain": "fine uncoated-paper fiber"
-  },
-  "elements": [
-    {
-      "what": "",
-      "role": "",
-      "motion": "",
-      "placement": ""
-    }
-  ],
-  "composition": {
-    "layout": "",
-    "negative_space": "",
-    "final_frame": ""
-  },
-  "motion_plan": "structure first, subject or cards second, action and result last",
-  "avoid": "typography, readable letters, numerals, logos, watermark, UI, subtitles, glossy 3D, photoreal environment"
-}
-
-### 色彩规则
-
-不要把 cobalt blue 当成唯一默认值。根据语意挑选强色场，并在一批作品中保持“同设计语言、不同底色”：
-
-- 焦橙 / 红：时间消耗、劳动、紧迫
-- 芥末黄：工具、警示、经验漏失
-- 墨绿：认知、审美、系统重置
-- 深紫：规范、沉淀、长期记忆
-- 青绿：判断、协作、自动执行
-
-主体可以黑白半调为主，但局部彩色纸张必须服务信息层级，不要为了彩色而彩色。
-
-### Imagegen prompt 模板
-
-使用 Codex 内置 `image_gen` 工具（如有本地 `imagegen` skill 则一并遵守）：
-
-```text
-Use case: ads-marketing
-Asset type: final still frame for a 9:16 image-to-video B-roll clip
-Primary request: Create a finished editorial paper-collage image expressing [一句话视觉命题].
-Scene/backdrop: perfectly flat [颜色] paper field [hex] with subtle uncoated paper fiber.
-Style/medium: premium editorial stop-motion paper collage; black-and-white halftone photographic cut-outs mixed with selective [点色] colored cardstock.
-Composition/framing: vertical 9:16 locked poster frame; central subject within the middle 70 percent; generous clean color-field negative space; 3–6 large separable paper groups for later assemble-from-empty animation.
-Materials/textures: visible printed halftone dots, crisp machine-cut edges, thin warm-cream paper keylines, soft low-opacity physical drop shadows.
-Constraints: [本条隐喻必须一眼看懂的关系].
-Avoid: no typography, no readable letters, no numerals, no logos, no watermark, no UI, no subtitles, no glossy 3D, no photoreal environment, no clutter.
-
-### 静帧 QA
-
-- 隐喻是否一眼看懂
-- 主体是否集中
-- 是否有假字、logo、水印或 UI
-- 是否保留足够纯色场，方便从空场组装
-- 是否是 3–6 个清晰大组，而不是满屏碎片
-- 同一批是否统一质感但有色彩变化
-
-将通过 QA 的原图复制到项目目录，生成带编号的静帧 contact sheet，展示给用户并停下等待 Gate 2 确认。静帧 QA 结论写入 `<project>/gate2-qa.md`。
-
-如果用户要求重生部分静帧，重生后生成 `still-contact-sheet-v2.jpg`（后续轮次递增 v3、v4…），保留旧版 contact sheet 不覆盖，方便对比。
-
-## Phase 3：用 Omni Flash 生成视频
-
-### 1. 准备首尾帧
-
-保留 imagegen 原图，再统一尾帧：
-
-```bash
--vf "scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920" \
-  <item>/frames/last-frame.png
-```
-```
-首帧默认是与尾帧相同底色的纯色空纸面：
-```
-```bash
--frames:v 1 <item>/frames/first-frame.png
-```
-```
-如果用户明确要求不从完全空白开始，首帧才保留一个基础物件。
-```
-
-### 2. 写 Omni 动画 prompt
-
-动作顺序默认采用：
-
-```text
-基础结构 → 人物或关键卡片 → 连接件 → 动作 → 最终结果
-```
-Prompt 模板：
-```
-```text
-Paper-collage stop-motion assembly, using Image 1 as the exact empty first frame and Image 2 as the exact completed last frame. In one continuous locked-off vertical shot, open on the empty flat [color] paper field.
-
-Assemble the scene piece by piece with crisp physical stop-motion timing: [按顺序描述 3–6 个元素如何滑入、卡位、连接和完成动作]. End by holding the supplied completed composition.
-
-Preserve the exact 9:16 framing, [hex] color field, colored cardstock accents, uncoated paper grain, halftone dots, cream keylines, crisp cut edges and soft shadows. Restrained tactile 2D paper craft only.
-
-No scene cuts, no camera movement, no zoom, no morphing, no new objects, no text, no letters, no numbers, no logos, no watermark, no UI, no sound.
-```
-每条 prompt 都要明确 Image 1 是空首帧、Image 2 是确认过的完成帧。最终构图必须贴近 Image 2，不让模型自由改造尾帧。
-```
-
-### 3. 检查 Omni 运行环境
-
-使用实际运行脚本的 Python 解释器检查 `google-genai` 版本。需要 `>= 2.10.0`。
-
-如果系统 Python 版本过旧或属于 externally managed environment，不要使用 `--break-system-packages`。使用共享隔离环境 `~/hyperframes-projects/.omni-venv/`——已存在就直接复用，只在不存在时创建，不要在每个项目内新建 `.venv`：
-
-```bash
-[ -x ~/hyperframes-projects/.omni-venv/bin/python ] || {
-  ~/hyperframes-projects/.omni-venv/bin/python -m pip install --upgrade "google-genai>=2.10.0" "httpx[socks]"
-}
-```
-```
-确认 `GEMINI_API_KEY` 已设置，但不要输出或记录密钥内容。
-```
-
-### 4. 批量调用 Gemini Omni Flash
-
-创建 `omni-jobs.json`。每个 job 使用两张图片做关键帧插值：
-
-```json
-{
-  "prompt": "<omni prompt>",
-  "image": [
-    "<item>/frames/first-frame.png",
-    "<item>/frames/last-frame.png"
-  ],
-  "output": "<item>/omni/run-v01/final-5s.mp4",
-  "aspect_ratio": "9:16",
-  "duration": 5
-}
-```
-使用本 skill 自带脚本：
-```
-```bash
-~/hyperframes-projects/.omni-venv/bin/python \
-  <本skill目录>/scripts/generate_video.py \
-  --batch <project>/omni-jobs.json \
-  --concurrency 3
-```
-```
-脚本默认模型即 `gemini-omni-flash-preview`。如果出现 legacy Interactions API schema 错误，说明用错了旧 SDK；切换到共享环境 `~/hyperframes-projects/.omni-venv/bin/python` 后重试，不要退回 Veo。
-```
-
-### 5. 强制无声交付
-
-即使 prompt 已写 `no sound`，仍用 ffmpeg 输出零音轨版本：
-
-```bash
--map 0:v:0 -c:v copy -an \
-  <run>/final-5s-noaudio.mp4
-```
-```
-默认交付 `final-5s-noaudio.mp4`，保留原始 `final-5s.mp4` 作为中间产物。
-```
-
-## 视频 QA
-
-不要只看尾帧，必须检查组装过程和最终落位。
-
-### Contact sheet
-
-```bash
--vf "fps=1,scale=270:480,tile=5x1" \
-  -frames:v 1 <run>/contact-sheet.jpg
-```
-```
-通过标准：
-
-- 首帧接近纯色空场；边缘轻微提前露出纸片可以接受
-- 中段能看到结构、人物或卡片逐步进入，而不是整体淡入
-- 没有切镜、zoom、3D 化或写实场景漂移
-- 没有假字、logo、水印或 UI
-- 最终帧与确认静帧一致；轻微姿态或细节漂移（如人物姿势微变、小零件增减）只要不影响隐喻语义即可判通过，不要为此重跑
-- 成片为 720×1280、24fps、5 秒、零音轨
-
-另外抽取视频末帧，与确认静帧并排生成 `end-frame-comparison.jpg`。批量项目再合并三张总览图：
-
-- `omni-contact-sheet-all.jpg`：全部成片逐秒抽帧
-- `video-first-frame-all.jpg`：全部成片实际首帧，验证真的从空色场开始
-- `end-frame-comparison-all.jpg`：确认静帧与视频末帧并排对照
-
-逐条 QA 结论（含带瑕疵通过的判定理由）写入 `<project>/gate3-qa.md`。
-```
-
-### 常见问题
-
-- 首帧边缘提前露出：轻微可接受；严格空场需求改用 HyperFrames 补前段
-- 组装感弱：缩短元素数量，并把 prompt 改为明确的逐件 slide in / snap into place 顺序
-- 尾帧漂移：强化 “Image 2 is the exact completed last frame” 和 “End by holding the supplied completed composition”
-- 出现假字：先回到静帧重生，不要直接用视频 prompt 修补
-- 个别视频失败：只重跑对应 job，不要重跑已经通过的条目
-
-## 默认交付
-
-向用户交付：
-
-- 每条 `<item>/omni/run-v01/final-5s-noaudio.mp4`
-- 每条 contact sheet
-- 批量总 contact sheet
-- 最终帧对照图
-- 一句说明每条文稿如何转成视觉隐喻
-
-如果成片问题来自 Omni 的快速生成限制，直接说明；只有需要精确图层控制时，才建议切换到 HyperFrames。
-
-## 旧 Veo 脚本
-
-目录中的 `scripts/generate_veo_first_last.py` 仅为旧项目兼容保留。不要在默认流程中调用它。只有用户明确要求 Veo 时才使用。
-
-## Source
-
-Rewritten for AICUT from [`pyang5166/gbro-collage-broll`](https://github.com/pyang5166/gbro-collage-broll)
-(MIT), SKILL.md. The method is that skill's own; the way it runs is AICUT's.