@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/jtxvideo-skill/SKILL.md

@@ -1,337 +0,0 @@
----
-name: jtxvideo-skill
-description: "JTX Video Skill: the method the original skill taught, run as an AICUT film — pages that draw any moment t, cut on tracks. Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "jackjls/jtxvideo-skill"
----
-
-# JTX Video Skill
-
-## 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.
-
-
-## Purpose
-
-Turn a recorded Chinese talking-head video into the user's standard AI-finance Douyin video:
-
-- the speaker is the visual anchor
-- motion graphics explain finance/AI/product events
-- the style is black-gold business, not flashy template editing
-- the workflow is staged and reviewable
-- final export includes a 2K 60fps version for publishing
-- the source orientation is respected: horizontal input stays horizontal unless the user asks otherwise
-
-## Non-Negotiable Workflow
-
-Never jump straight from source video to a finished edit. Move in this order:
-
-1. **Create project folder**
-   - Use `hyperframes-runs/<project-slug>/`.
-   - Keep all generated files inside the project folder.
-
-2. **Preprocess source video**
-   - Convert the input video to stable H.264 at 30fps with fixed keyframes.
-   - Default command shape:
-     ```
-- Save as `workflow/01_preprocessed_keyed.mp4` and copy/use as `assets/talk-main-keyed.mp4`.
-   - Immediately run `ffprobe` and record source orientation:
-     - `width > height` = horizontal workflow
-     - `height > width` = portrait workflow
-   - Save the orientation decision in the project notes / `design.md`. Do not choose layout rules before this decision is made.
-
-3. **Extract audio and transcribe**
-   - Extract clean audio, usually `workflow/01_audio_16k.wav`.
-   - Use a Chinese-capable high-accuracy transcription model. Do not use `.en` Whisper models for Chinese.
-   - First output to the user must be the SRT file path and content summary only.
-   - Save SRT as `workflow/01_audio_16k.srt`.
-
-4. **SRT 文件确认**
-   - Do not analyze the script, storyboard, or render before the user confirms SRT.
-
-5. **Analyze content structure**
-   - Segment the transcript into content types:
-     - 开场钩子
-     - 痛点呈现
-     - 正误对比
-     - 流程清单
-     - 方法论讲解
-     - 案例 / 数据
-     - 行动号召
-     - 金融观点 / 市场判断
-     - AI工具实测 / 产品事件
-   - Save as `workflow/05_content_structure_for_review.md`.
-   - Ask the user to review.
-
-6. **文案结构确认**
-   - Do not generate storyboard until confirmed.
-
-7. **Generate storyboard**
-   - Use `templates/storyboard-template.md` as the schema.
-   - Include time range,口播内容,文案类型,画面策略,动效元素,音效建议,审核关注点.
-   - Before writing the storyboard, branch by source orientation:
-     - Horizontal: material-first split layouts are allowed for long evidence/chart explanations.
-     - Portrait: the speaker must remain on screen; materials and motion graphics attach as overlays above the speaker.
-   - Save as `workflow/06_storyboard_for_review.md`.
-   - Ask the user to review.
-
-8. **分镜表确认**
-   - Do not generate HyperFrames code before storyboard confirmation.
-
-9. **设计确认并生成项目 design.md**
-   - Copy `templates/design-default.md` into the project as `design.md`.
-   - Adapt only what the current video needs.
-   - Keep the approved defaults unless the user asks otherwise.
-
-10. **Generate HyperFrames project**
-    - Use one continuous talking-head video as the base layer.
-    - Do not repeatedly mount many slices of the same MP4; it can cause gray preview, freezing, or the speaker disappearing.
-    - Put captions and visual explanations in separate timed overlay layers.
-    - Prefer concise dynamic keywords over dense cards when there is no external material to show.
-    - Use real source images/charts when they are available and relevant; do not redraw unclear symbolic diagrams.
-    - Apply the orientation-specific layout rules:
-      - Horizontal: may use left-material / right-speaker split scenes when materials need sustained explanation.
-      - Portrait: never switch to a full-screen material scene; the speaker stays visible and materials/dynamic text are overlay layers attached to the speaker shot.
-    - Main video must be muted; use a separate `<audio>` track.
-
-11. **最终视觉确认**
-    - Generate preview/review renders.
-    - Use screenshots/contact sheets for visual checks.
-    - Apply user feedback on subtitle height, size, transparency, person brightness, card placement, and motion restraint.
-
-12. **Final export**
-    - Preserve the source orientation unless the user explicitly asks to convert.
-    - Horizontal final version: `2560x1440 / 60fps`.
-    - Portrait final version: `1440x2560 / 60fps`.
-    - Render a high-quality master, then generate the final 2K 60fps publishing file.
-    - Re-mux full original audio after HyperFrames rendering.
-    - Verify with `ffprobe`.
-```
-
-## Editing Rules Learned From Review
-
-Use these rules before writing the storyboard and again before rendering:
-
-1. **Find the real topic first**
-   - Do not animate every sentence.
-   - Extract only the few beats that carry the video: event, key data, core contrast, bottleneck, conclusion.
-   - For industry-chain videos, prioritize the event trigger, market-size numbers, margin/comparison numbers, and the true bottleneck.
-
-2. **Use fewer, sharper words**
-   - If there is no source material, show only dynamic keywords or numbers.
-   - Do not use generic cards just to fill the screen.
-   - A key overlay should usually be one line, one number, or one contrast.
-   - Avoid stacking many explanatory lines; if everything is highlighted, nothing is highlighted.
-
-3. **Use source material as the main visual when available**
-   - Horizontal: if a chart/table/image is the evidence, make the material readable and visually dominant.
-   - Portrait: source material can be shown, but only as an overlay on top of the speaker shot unless the user explicitly asks for a full-screen insert.
-   - Do not crop important parts of the material.
-   - Do not cover the material with duplicate text or redundant labels.
-   - Put only the key prompt above or near the material.
-
-4. **Decide horizontal vs portrait before layout**
-   - Always inspect `width` and `height` before storyboard/design/code.
-   - Do not reuse horizontal split-screen logic on portrait footage.
-   - Do not reuse portrait overlay logic when a horizontal video needs readable long-form evidence.
-   - Record the decision in `design.md` with the chosen canvas and export target.
-
-5. **Portrait layout rules**
-   - The speaker must remain visible for the entire video.
-   - Materials, charts, screenshots, and keywords are overlay explanations attached to the speaker shot.
-   - Do not create full-screen material takeovers in portrait unless the user explicitly asks.
-   - If a source image is provided for a portrait video, use it briefly as an upper overlay, side overlay, or small evidence strip; do not let it replace the speaker.
-   - Do not cover eyes, mouth, or key facial expression areas.
-   - Prefer top/upper-side overlays for keywords and data; keep bottom reserved for subtitles and Douyin-safe UI.
-   - If a phrase's visual emphasis feels like a distracting popup, remove that emphasis and keep the phrase in the full subtitle only.
-
-6. **Horizontal split layout for long material explanation**
-   - For horizontal videos with important material, use a two-panel layout:
-     - left: large material panel
-     - right: synchronized speaker panel
-   - Keep material and speaker panels aligned in height.
-   - Leave enough gap between title, panels, and subtitles.
-   - The material is the primary visual in these sections; the speaker supports the explanation.
-
-7. **Static blurred background in split scenes**
-   - When the speaker is shown in a right-side panel, the background behind the panels must not be a live moving copy of the speaker.
-   - Use a still frame from the source video as a blurred, darkened background.
-   - Only the speaker panel should move.
-   - This prevents the viewer from seeing two moving speakers at once.
-
-8. **Animation must match the spoken timing**
-   - Do not reveal a number or conclusion before the matching subtitle/口播.
-   - Tie overlays to the exact SRT segment that says the key phrase.
-   - Example: `$40 -> $60` should appear when the speaker says "从40美金涨到60美金", not when earlier context begins.
-
-9. **Text style**
-   - Use solid, heavy, high-contrast Chinese typography for key statements.
-   - Avoid hollow, striped, folded, decorative, or textured type unless explicitly requested.
-   - Use gold only for the most important number or conclusion.
-
-10. **Subtitles stay independent**
-   - Subtitle bars must not overlap material panels or speaker panels.
-   - In horizontal split scenes, reserve a clear bottom subtitle band below the main panel.
-   - In portrait scenes, subtitles remain the base layer of comprehension; overlays must not compete with the subtitle bar.
-   - Subtitle background should be translucent, not fully opaque black.
-
-11. **Render stability**
-   - Use one continuous main video as the base.
-   - For picture-in-picture speaker panels, prefer pre-cut speaker clips as root-level media layers controlled by opacity.
-   - Avoid dynamically creating nested video elements inside timed HTML scenes.
-   - Prefer local bundled scripts such as `assets/gsap.min.js` over CDN scripts for renders.
-   - After rendering, re-mux full original audio and verify that audio duration is close to video duration.
-
-## Default File Contract
-
-Use this structure:
-
-```text
-hyperframes-runs/<project-slug>/
-  assets/
-    talk-main-keyed.mp4
-    talk-main-audio.m4a
-  workflow/
-    01_preprocessed_keyed.mp4
-    01_audio_16k.wav
-    01_audio_16k.srt
-    05_content_structure_for_review.md
-    06_storyboard_for_review.md
-  templates/            # optional project-local copies
-  design.md
-  index.html
-  renders/
-    review.mp4
-    contact-sheet.jpg
-    preview-frame.jpg
-    final-2k60.mp4
-
-## Visual Defaults
-
-Use `templates/design-default.md` as the source of truth for:
-
-- AI金融黑金商务风
-- transparent, readable subtitle bar
-- bright speaker/person layer
-- concise dynamic keywords and restrained motion
-- material-first split layout when charts/tables/images are important
-- portrait overlay layout when the source video is vertical
-- no noisy gold line or flashy decoration
-- Douyin-safe subtitles
-
-## Quality Gates
-
-Before telling the user a render is ready:
-
-- Run `npx hyperframes validate`.
-- Run `npx hyperframes inspect --samples 12` or denser for risky layouts.
-- Use `npx hyperframes snapshot --at ... --describe false` for exact timestamps that received feedback.
-- Run `ffprobe` on final MP4.
-- Confirm video stream dimensions, frame rate, and duration.
-- Confirm audio duration is complete and close to video duration.
-- Generate at least one preview frame or contact sheet.
-- For split scenes, inspect a frame while the section is active and confirm the blurred background is static.
-- For portrait scenes, inspect frames where materials appear and confirm the speaker remains visible.
-- For timed key numbers, inspect one frame before the cue and one frame after the cue.
-- For any user-reported flicker/popup issue, extract several frames around the timestamp from the rendered MP4, not only from HyperFrames snapshots.
-
-## Export Defaults
-
-Review version:
-
-- Match source orientation unless the user asks otherwise.
-- Horizontal review: `2560x1440` or a fast proxy at the same aspect ratio.
-- Portrait review: `1080x1920`.
-- `30fps` or `60fps` depending on iteration speed
-- complete audio remuxed from the original processed audio
-
-Final publishing version:
-
-- Horizontal: `2560x1440`
-- Portrait: `1440x2560`
-- `60fps`
-- H.264 High Profile
-- AAC audio
-- `+faststart`
-
-Recommended horizontal final encode shape:
-
-```bash
--map 0:v:0 -map 1:a:0 \
-  -vf "fps=60" \
-  -c:v libx264 -preset slow -crf 17 -pix_fmt yuv420p \
-  -profile:v high -level 5.2 -movflags +faststart \
-  -c:a aac -b:a 192k final-horizontal-2k60.mp4
-```
-```
-Recommended portrait final encode shape:
-```
-```bash
--vf "scale=1440:2560:flags=lanczos,fps=60" \
-  -c:v libx264 -preset slow -crf 17 -pix_fmt yuv420p \
-  -profile:v high -level 5.2 -movflags +faststart \
-  -c:a copy final-2k60.mp4
-```
-
-## Avoid
-
-- Do not cut before SRT confirmation.
-- Do not skip storyboard review.
-- Do not make the production process itself appear in the video unless that is the topic.
-- Do not add full-height scrolling gold lines or flashy decorative effects.
-- Do not darken the entire speaker with a heavy black overlay.
-- Do not put subtitles too low where Douyin UI covers them.
-- Do not make subtitle backgrounds fully opaque black.
-- Do not convert horizontal source video to portrait unless the user explicitly asks.
-- Do not convert portrait source video to horizontal unless the user explicitly asks.
-- Do not use dense cards when a short dynamic keyword or real material is enough.
-- Do not show a live moving background speaker behind a picture-in-picture speaker panel.
-- Do not let a portrait video become a full-screen PPT/material video; the speaker must stay visible.
-- Do not reveal key numbers before the matching口播/SRT timing.
-- Do not invent financial facts, stock claims, prices, dates, or performance data.
-- Do not imply investment advice or喊单.
-
-## Response Style
-
-When running this skill, report the current stage and next approval point. Keep the user in control of:
-
-- SRT confirmation
-- 文案结构确认
-- 分镜表确认
-- 设计确认
-- 最终视觉确认
-
-If the user says "继续", continue to the next unblocked stage.
-
-## Source
-
-Rewritten for AICUT from [`jackjls/jtxvideo-skill`](https://github.com/jackjls/jtxvideo-skill)
-(MIT), skills/jtxvideo-skill/SKILL.md. The method is that skill's own; the way it runs is AICUT's.

package/skills/kinetic-reel/SKILL.md

@@ -1,140 +0,0 @@
----
-name: kinetic-reel
-description: "Make kinetic-typography motion reels (MP4) in code: showreels, portfolio or work reels, product promos, intro films, \"motion design\"-style videos with bold condensed type, HUD micro-type, black/cream/lime palette, three.js generative layers (particle terrain, liquid marble, chrome knot, particle clouds), shape-continuity transitions and a synthesized, beat-locked score. Canvas 2D + WebGL rendered frame by frame in headless Chrome, encoded with ffmpeg. Use when the user asks for a reel, showreel, motion-graphics / kinetic-type video, a \"Work Reel\" in this style, or to extend/re-cut the Work Reel ’26. Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "tuzhechen2005/opus-video-skills"
----
-
-# Kinetic reel
-
-## 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.
-
-
-A method for making editorial, high-energy motion reels entirely in code, distilled from "Work Reel ’26" (an 84 s
-portfolio reel: Microsoft → UW–Madison Surgery → Enterprise RAG → trace finale). It's bundled as a runnable
-`template/` (a 17 s demo that uses every technique) and a full worked example in `examples/work-reel-26/`.
-
-**How it works.** Every frame is a pure function of `t`. `reel/index.html` loads `reel/cues.js` (the timeline) and
-`reel/reel.js` (an ES module: three.js layers, a 2D type canvas and a WebGL post pass). `render.mjs` drives headless
-Chrome to render contact sheets, strips and stills for review, or every frame in parallel, and ffmpeg encodes the MP4.
-`music/score.mjs` synthesizes the score in Node from the same `cues.js`, so every hit lands on its frame.
-
-## Workflow
-
-### 0. Scaffold
-```bash
-bash ~/.claude/skills/kinetic-reel/scripts/new_project.sh <project-dir>
-```
-```
-Then read [references/style-guide.md](references/style-guide.md) (the look, the motion vocabulary, the transition
-catalogue, layout rules) and [references/sound.md](references/sound.md) in full before designing. Skim
-`examples/work-reel-26/reel.js` for how a long reel is organised.
-```
-
-### 1. Decide the emphasis, then storyboard
-- Ask (or settle) **what leads and how much time each part gets**. Viewers notice proportions: v2 of the Work Reel had
-  Microsoft at ~60% and had to be rebalanced. The fix was to *add* runtime for the other chapters rather than cut the
-  one people loved. Aim for the lead ≈ 35–40%.
-- Write `STORYBOARD.md` from the template: time budget, the source of truth for facts, a unifying device, and one row
-  per shot (headline, visual mechanism, transition in, sound).
-- **Facts are sacred.** Only figures from the provided source, each with its scope ("~180 self-built cases",
-  "320 bilingual prompts"). Anything illustrative (tool IDs, doc IDs, example answers, an AST) is labelled
-  `SCHEMATIC` in the HUD. Never draw per-dimension data you don't have. Visualise the *process*. Medical/clinical
-  work: no efficacy claims, say "in development / not yet validated" when true.
-
-### 2. Build
-- `reel/cues.js`: `S` (scene starts), `hits` (every beat that matters), `CH` (chapters → the segmented progress bar),
-  `dur`. 120 BPM: cuts on beats, chapter changes on bars.
-- `reel/reel.js`: one function per shot, `sName(t)`. Each paints the whole frame: `bg()` → GL layer via
-  `glShot(...)` → type and shapes → `hud(t, {...})`. Register the shot in `SCENES` with a key; add a transition into it
-  in `TR` (or leave it out for a hard cut on the beat).
-- Keep the `CJK` constant listing every Chinese character you draw (fonts load as a subset).
-
-### 3. Look at it, every shot, several times
-```bash
-node render.mjs --sheet=1,4,7,10,13,15.5 --cols=3 --w=640 --out=out/check/sheet.jpg   # key frames
-node render.mjs --stills=10.4 --out=out/stills                                          # full-res detail
-```
-```
-Open every image with Read. Check for collisions (labels over lines, stats over stats, text into HUD bands), empty
-frames on a cut, legibility at 1600 px wide, and that each transition *starts on something* and lands cleanly.
-Budget: a sheet per chapter, a full-res still per dense shot, a strip per transition type.
-```
-
-### 4. Score, render, encode
-```bash
-node music/score.mjs && ffmpeg -y -i music/score.wav -af "highpass=f=30,equalizer=f=250:t=q:w=1:g=-2,loudnorm=I=-14:TP=-1.2:LRA=8" -c:a aac -b:a 192k assets/score.m4a
-node render.mjs --frames --workers=4                                    # ~8–40 ms/frame on a Mac GPU: seconds, not minutes
-node render.mjs --encode --audio=assets/score.m4a --out=out/reel.mp4 --fps=30
-```
-
-# web copy: ffmpeg -i out/reel.mp4 -vf scale=1600:900:flags=lanczos -c:v libx264 -preset slow -crf 23 -c:a aac -b:a 128k -movflags +faststart out/reel-web.mp4
-```
-You can't hear the score: check it with `showspectrumpic` and per-section RMS, and say so when you report. Audio-only
-changes don't need a re-render: remux (`-map 0:v -map 1:a -c:v copy`).
-```
-
-## The rules that matter most
-1. **Type is the hero.** Condensed display (Anton) for headlines, one wide grotesk (Archivo Black) for slams, italic
-   serif (Instrument Serif) for the one "voice" line, mono (JetBrains Mono) for every label. Big English headlines,
-   small Chinese annotations when the audience is bilingual.
-2. **Every shot has a mechanism**, not just words: a grid routing, a matrix scanning, a token hopping a state machine,
-   lists fusing, packets passing gates. The headline names it; the mechanism proves it.
-3. **Beat-locked.** Letters, chips, counters and cuts sit on `hits`. Fast in (expo-out), hold, next.
-4. **Continuity.** Prefer shape-continuity transitions (zoom through the element the eye is on, iris from the thing
-   that just lit up) to hard cuts inside a chapter. Hard cuts + an impact are for chapter slams.
-5. **No noise whooshes.** In-key swells into chapters, quiet tonal cues inside them, silence on about half the cuts.
-6. **Honest data** (see step 1) and a `SCHEMATIC` label on anything illustrative.
-7. **Guard the math.** Radii from `backOut()` can dip below 0 (canvas throws): `circle`/`starburst`/`jelly` already guard.
-   In the synth, clamp local time to ≥ 0 before `pow()` and clamp SVF `f`/`q` (both already in the template).
-
-## Shipping a reel as a site intro
-The Work Reel ships as its site's entry page: `index.html` plays the reel full-screen and hands off to the portfolio.
-The pattern is bundled in [examples/site-intro/`site-intro` (`index.html`, `css/intro-reel.css`,
-`js/intro-reel.js`, plus a Cloudflare Worker that adds HTTP Range support):
-- Try to autoplay with sound. If refused, play muted with a pulsing "tap for sound". If that is refused too, or
-  `prefers-reduced-motion`, show the end screen in "start" mode (play is the lead button).
-- During playback: small glass pills at top centre, since the reel's HUD owns the corners. "Skip intro" carries a
-  ring that fills with the reel. A sound toggle's equalizer bars pulse at the reel's BPM. Esc skips, M mutes.
-- End screen: the reel's grammar in HTML. Corner brackets, the name as drop-in letters (Anton), mono micro-type,
-  a segmented trace bar, a 2D dot-field version of the particle terrain, a lime primary CTA (text roll on hover,
-  sheen, magnetic, a pulse every bar), and a ghost "watch again".
-- Leave through a lime iris from the pressed control, duck the audio, then navigate.
-- Two encodes (1600×900 at ~1.3 Mbps, 960×540 at ~0.6 Mbps for phones / save-data), `preload="auto"`, a poster.
-  Serve them with HTTP Range (206), or seeking and Safari break.
-
-## Source
-
-Rewritten for AICUT from [`tuzhechen2005/opus-video-skills`](https://github.com/tuzhechen2005/opus-video-skills)
-(MIT), skills/kinetic-reel/SKILL.md. The method is that skill's own; the way it runs is AICUT's.
-
-The repository holds 2 skills; this is the one its entry stands for.

package/skills/kinetic-reel/examples/work-reel-26/STORYBOARD.md

@@ -1,51 +0,0 @@
-# Work Reel ’26 · 动态排版版
-
-**当前版本 v3**：84 s，16:9，30 fps，120 BPM。v1（42 s）和 v2（52 s，微软约占 60%）的源码存档在 `kinetic/v1/`、`kinetic/v2/`，
-对应视频是 `out/kinetic-v1*.mp4` 和 `out/kinetic-v2*.mp4`。
-
-## v3 结构（按重点排序：微软 → UW 外科系 → 企业 RAG）
-
-| 时间 | 段落 | 画面 |
-|---|---|---|
-| 0–10 | 开场 | 同 v1：BOOT → EVERY ANSWER NEEDS EVIDENCE → 名字 |
-| 10–12 | 目录 | 01 微软 / 02 UW–Madison 外科系 / 03 企业 RAG，各带角色和时长；镜头从「01」钻进下一镜 |
-| 12–42 | **01 微软**（30 s） | 标题 → 24 个工具格 → 6 维评测矩阵 → 约束解码 → 61%→87% → ReAct 状态机 → AutoGen 规划/审查 → 事件管道 → FAIL SAFE → 回顾胶带 |
-| 42–60 | **02 UW–Madison 外科系**（18 s） | DEPARTMENT OF SURGERY（字母里填流体纹理）+ 粒子聚合成甲状腺 + **AI RESEARCHER** → 病历切块进入向量星团，查询检索 top-k → 每句话都有出处 → 四道安全闸门（一个答案被拒答并转交医生）→ *Cite it — or abstain.* |
-| 60–74 | **03 企业 RAG**（14 s） | 标题配铬金属扭结 → BM25 与稠密两路召回、RRF 融合（按真实 RRF 公式排序）+ 级联重排漏斗 → DistilBERT 意图路由加 LLM 兜底 → SQL AST 守卫拦截 DROP → 高并发指标 |
-| 74–78 | Trace | 整支片展开成一次 Agent 运行的 span 瀑布图，STATUS OK |
-| 78–84 | 片尾 | 名字、水母标志、联系方式 → 故障效果 → 收成一个点 |
-
-**v3 新增的手法**：镜头之间不硬切，改用形状连贯的转场，包括钻进被选中的 UI 元素、圆形放大、扫描线擦除、百叶窗、像素格溶解、推移。
-底部进度条按章节分段，整支片就是一条 trace。三维粒子云从灰尘聚合成甲状腺，再散开成向量空间，检索线接到按三维投影算出的真实点上。
-
----
-
-## v1 原始分镜（42 s，供参考）
-
-参考风格：一段动态排版参考片（未收录）。粗体窄字号大标题加 HUD 式小字标注；配色是黑、米白、荧光青柠，点缀电光蓝和信号红；
-配合生成式画面（粒子地形、流体大理石纹、铬金属扭结、方框隧道、数据柱）和卡拍转场（斜切、切片、RGB 分离）。
-纸海里的 Jelly 在这里简化成一个几何水母标志，顶替参考片里星芒的位置。
-
-**文字规则**：大字用英文，小字标注里夹少量中文。所有数字都来自简历和网站原文，附带口径说明（测试集、样本量）。
-医疗相关只写「决策支持」「研发中、尚未进行患者端验证」。
-
-**节拍**：120 BPM，一拍 0.5 s，一小节 2 s，每次切镜都卡在拍点上。配乐是 `music/score_kinetic.mjs`，读取 `kinetic/cues.js`。
-
-| 时间 | 段落 | 画面 |
-|---|---|---|
-| 0–4 | BOOT | 黑底，HUD 四角括号画出，时间码开始走；粒子地形升起 → 2.0 中心亮起青柠圆点，卡拍跳动 → 3.0 展开成水母标志 → 闪白 |
-| 4–6 | EVERY / ANSWER | 巨大的 EVERY 加上下描边叠字滚动 → 5.0 翻转成 ANSWER |
-| 6–7 | needs | 米白底，细圆圈画出，衬线斜体 *needs*，一个红点 |
-| 7–8 | EVIDENCE | 电光蓝满屏，EVIDENCE 带动态模糊砸进来；中文小字「每个回答，都需要依据」；切片故障出场 |
-| 8–10 | 名片 | ZHECHEN TU. 从描边到实心，RGB 故障入场；AI APPLICATION · AGENT ENGINEER，涂喆宸，衬线副标 |
-| 10–13 | 01 企业 RAG | 米白底，ROUTE. / RETRIEVE. / RERANK. 逐字母落下；青柠圆里的水母标志；每行配一个技术标签 |
-| 13–16 | 01 数据 | 黑底，流体纹理填充的柱状图加红色曲线；计数器滚到 093%（路由准确率），然后 +18% Recall@10；S 曲线面板 → 斜切转场 |
-| 16–19 | 02 多智能体 | 旋转的青柠/白色方框隧道，中心圆内写 FIVE / AGENTS / ONE CASE，五个角色标签卡拍绕圆出现 |
-| 19–20 | 02 红旗 | 信号红满屏 RED FLAG，画面震动，写明「转交医生」 |
-| 20–22 | 02 指标 | 96%+ 急症召回率，配三个小指标；口径：约 180 条自建用例，决策支持、不做诊断 |
-| 22–25 | 03 微软 | 青柠底 CALL / THE RIGHT / TOOL，右侧铬金属扭结旋转；24 个 Azure REST 工具 |
-| 25–28 | 03 指标 | 61% → 87% 像里程表一样翻滚（Phi-3 完整调用准确率，320 条双语评测） |
-| 28–30 | 04 甲状腺 | 流体大理石纹满屏，衬线斜体 *Cite it — or abstain.* 逐词出现 → 竖条切片 |
-| 30–32 | 04 信息 | 米白底 CLINICAL RAG 大字；右列逐条打勾：主导 AI 工程师、6.9 万美元资助、两位 PI、依据生成、安全评测 |
-| 32–36 | 04 胶带 | 两条斜向胶带反向滚动：TRACEABLE ✳ GROUNDED ✳ SAFE / CITE ↗ SOURCE ↗ ABSTAIN；巨大的描边 04；底部小字「研发中，尚未进行患者端验证」 |
-| 36–42 | 片尾 | ZHECHEN / TU 字母升起，青柠水母标志旋转，圆形进度环；邮箱和 GitHub → 40.5 RGB 故障 → 收成一个点 → 黑场。片尾只用英文 |

package/skills/kinetic-reel/references/sound.md

@@ -1,40 +0,0 @@
-# Sound: the synthesized, beat-locked score
-
-`music/score.mjs` is a small offline synth in Node (no samples, no dependencies). It reads `reel/cues.js`, places
-every sound in time with `put(t0, len, gen(tt), {gain, pan, rev, bus})`, and writes `music/score.wav`.
-There are two buses. `drum` stays dry. `mus` is ducked by the kick (sidechain): each `kick()` writes into `duck[]`.
-A Freeverb send, tanh glue and normalization follow; loudnorm (−14 LUFS) runs in ffmpeg.
-
-## Instruments
-`kick`, `clap`, `hat(open?)`, `tick` · `bass` (saw/square through tanh) · `pluck` · `pad` (detuned, slow swell) ·
-`stab` (chord hit) · `blip` (in-key sine ping) · `odo` (a burst of ticks for a rolling counter) · `glitch`
-(square-wave crumbs) · `impact` (kick + low-passed noise tail + sub drop) · `softImpact` (no noise tail) ·
-`riser` (tonal sweep with a little low-passed noise).
-
-## Transition sounds (a hard rule, learned from feedback)
-**No noise whooshes on transitions.** v3 fired 29 in 84 s; it was "cool at first, then annoying". Use:
-- `swell(tEnd, chordNotes, len)`: a reversed pad that grows and stops dead on a chapter cut. Match the chord.
-- `bloom(t, notes)`: a soft bell chord, for iris transitions.
-- `glide(t, len, m0, m1)`: a quiet triangle glide up, for a zoom-through.
-- `harp(t, span, notes, v, dir)`: a quick pluck run, for slats (vary the direction).
-- `hatRoll(t)`: 3–4 rising hats, for wipes.
-- **nothing**, for about half the cuts: the downbeat and the next shot's own entrance sounds (tile ticks, scan blips,
-  typing) carry them.
-Never use the same cue twice in a row. Keep `impact()` (with its noise tail) for about five structural moments. Use
-`softImpact()` for other slams.
-
-## Arrangement
-- A groove loop runs over the reel: 4-on-the-floor kick, claps on 2 and 4, 8th hats (16ths for the lead chapter),
-  a two-note bass per beat, a pluck arpeggio, and a pad per bar. Chords: Am F C G.
-- Give chapters different energy. A serious chapter goes half-time: kick on beat 1, soft clap on 3, a *decaying* sine
-  sub (not a sustained drone). Aim for about 5 dB RMS lower than the lead chapter. Check with per-section RMS.
-- Take the drums out under a glitch exit. Leave a breath (kick only) on a quiet word.
-- Tie hits to the picture: ticks per letter/tile/row, blips on chips and nodes, `odo` under counters, a stab when a
-  number lands, a low buzz on a rejection, a bell per retrieved item.
-
-## Gotchas (both cost real time)
-- `t0 * SR` isn't an integer, so the first sample's local time can be −1e−16, and `Math.pow(negative, 2.2)` is NaN,
-  which silences the whole mix after normalization. `put()` clamps `tt >= 0`. Keep it.
-- The state-variable filter in `noise()` explodes near `f = 1`: `f` is clamped to .85 and `q` to ≥ .45. Keep it.
-- You can't listen. Verify with `ffmpeg -lavfi showspectrumpic`, per-section RMS, and a count of the calls you
-  meant to remove (e.g. `grep -c "whoosh("`). Tell the user you couldn't hear it.

package/skills/kinetic-reel/references/style-guide.md

@@ -1,79 +0,0 @@
-# Style guide: the Work Reel look
-
-## Palette (`K` in reel.js)
-| token | hex | use |
-|---|---|---|
-| `ink` | `#0E0F0E` | default dark ground, type on light grounds |
-| `ink2` | `#171917` | nodes/cards on ink |
-| `cream` | `#F1EEE6` | light ground, type on dark |
-| `lime` | `#DDF53D` | THE accent: the lead chapter, the chosen item, the "OK" state, the mark |
-| `blue` | `#2F3CFF` | second ground (one slam per reel, one full-bleed scene), secondary emphasis on cream |
-| `red` | `#E8412F` | danger/rejection only (a red flag, a blocked query, a failed gate) and the full stop in a headline |
-| `grey`/`mid` | `#8B908A` / `#5E625D` | dimmed type |
-Rotate grounds between shots (ink → lime → ink → cream → blue …) so every cut is a visible change. A darker
-`#0B0D0C` ground marks a more serious chapter.
-
-## Typography (`F` in reel.js)
-- Headlines: `font(F.cond, 130–330)`, one or two lines, ending in a full stop. Colour the key word (lime on dark, blue
-  on cream) or just the full stop (red).
-- Slams: `font(F.wide, 172–250)` (Archivo Black) for one-word statements (EVERY / ANSWER / EVIDENCE), with outlined
-  echoes stacked above and below.
-- Voice line: `italic 124px ${F.serif}` over the liquid layer, word by word on beats.
-- Labels, HUD, chips, tables: `font(F.mono, 14–26, 700)` with 2–8 px letter-spacing. Chinese: `font(F.zh, 20–34, 500)`
-  with 4–14 px spacing.
-- Numbers: `F.cond` at 84–420 px. Counters roll (`count()` expo) or flip (`odometer()`).
-
-## Layout
-- Canvas 1920×1080. Left margin **140**, right edge **1780**. Headline baselines around y 250–470.
-- The HUD owns the four corners: brackets at 40 px, micro-type at (58, 63) / (1862, 63) / (58, 1022) / (1862, 1022),
-  one optional Chinese line at y ≈ 988 (left or right), the segmented progress bar at y 1050. Keep content out of
-  y > 950.
-- Put stats in a row at y ≈ 900 or stacked in a right column at x ≈ 1400–1540. Never two stats in the same 300 px
-  horizontal band at the same height: they collide. Check with a full-res still.
-
-## Motion vocabulary (all in reel.js)
-| helper | what it does |
-|---|---|
-| `dropWord(s, x, y, font, t0, t, {fall, stag, fillAt})` | letters fall in with expo-out and a stagger; `fillAt(i)` recolours one letter |
-| `letters(s, x, y, o, fn)` | per-letter transform (dy, r, s, a, fill): build your own entrances |
-| `chip(s, x, y, {k, bg, fg})` | a label that wipes in; returns its width so chips can be laid out in a row |
-| `stat(v, label, x, y, {k, size, col})` | a big number with its label |
-| `count(t, t0, t1, v0, v1)`, `odometer(v0, v1, k, x, y, font, col)` | number roll / per-digit flip |
-| `arrow`, `line`, `circle`, `rrect`, `starburst`, `asterisk`, `arrowNE`, `jelly` (a simple mark) | shapes |
-| `redacted(x, y, w, h, seed, col)` | "text" as rounded blocks: never invent real content |
-| `thread(p0, p1, k, col)` | a cubic "citation" line drawn up to k |
-| `texWindow(draw, src)` | letters as windows onto a shader layer (e.g. a headline filled with liquid marble) |
-| `tape(t, y, rot, col, fg, words, speed, dir, k)` | a diagonal scrolling tape with ✳ / ↗ separators |
-| `glitchIn(lt, seed)` | slice + RGB split on a hard cut in (auto-disabled during transitions) |
-| `hud(t, {tl, tr, bl, br, zh, dark, brackets})` | the corner system + segmented progress bar |
-
-## GL layers (render into `glA`, then `glShot()` draws it into the 2D frame)
-- `terrain(t, {fade, amp, push, camY})`: additive particle landscape, for openings and chapter titles on ink.
-- `liquid(t, {zoom, seed, strips, stripAmt})`: blue/red/lilac domain-warped marble, for the one "principle" line. With
-  `strips` it becomes a vertical-slice glitch for exits.
-- `knot(t, {x, y, s})`: a chrome torus knot with an inked outline, for "form"/"tool" shots on lime or cream. Shade it
-  with a soft oval shadow.
-- `cloud(t, {k, rot, x, y, s, hi})`: 9k particles morphing from a shell (k 0) to a silhouette (k 1).
-  `cloud.project(i, t, o)` returns a point's screen position, so 2D lines can hit real 3D points (retrieval).
-  Swap the silhouette's `inside(x, y)` for your subject.
-- Post pass (`FX`): `split` (RGB offset), `slice` (row glitch), `grain`, `vign`, `flash`.
-
-## Transitions (the `TR` map: `key: [duration, fn]`)
-| fn | use it when | example |
-|---|---|---|
-| `trZoom(rect or () => rect)` | the eye is on a UI element (a tile, a chip, a card): zoom *through* it into the next shot | tile T07 → the eval matrix; the "01" in the index → the chapter title |
-| `trIris(cx, cy, ring)` | something just lit up / a dot / a node: the next shot grows out of it | the ANSWER node → the next scene |
-| `trPush(dx, dy)` | a change of topic at the same level; add nothing else | diagram → diagram |
-| `trWipe(col)` | a scan / sweep already moving across the screen | the matrix scan line → the next shot |
-| `trSlats(n)` | a "reveal" beat, rhythmically, on a bar | pipeline → pipeline |
-| `trGrid(cols, rows, col)` | ending a section on a grid-like image (tests, pixels) | the test grid → the recap tapes |
-Scenes register the rect they'll be zoomed through in `RECT` while they draw (A is rendered frozen on its last frame,
-then B live; A's FX are dropped). Keep each transition 0.45–0.6 s and vary them: never three of one kind in a row.
-
-## Recurring devices that work
-- An **index/contents** card that states the order of emphasis (with each chapter's role and duration).
-- The **trace** device: a segmented progress bar per chapter, and a finale that lays the whole reel out as a span
-  waterfall ("STATUS OK").
-- A **mark** that recurs (start: a dot becomes the mark; end: the mark next to the name; the reel collapses to a dot).
-- Tapes in two directions for a chapter recap.
-- Red for exactly the failures that the system catches (red flag, blocked DROP, failed gate).