@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/package-talking-head-video/docs/superpowers/specs/2026-08-08-multicolor-center-speaker-design.md

@@ -1,154 +0,0 @@
-# Multicolor Center-Speaker Packaging Design
-
-## Status
-
-Approved by Nana on 2026-08-08. Implementation may proceed only against this specification. The public project must describe its capabilities directly and must not discuss private tooling, omitted systems, or hidden implementation sources.
-
-## Problem
-
-The current renderer advertises eighteen structures, but every structure passes through one dark rounded `StructureShell`. The result changes labels and minor decorations while preserving the same panel silhouette, typography hierarchy, shadow, and single cyan accent. The real-case preview therefore looks like one card repeated across the whole video.
-
-The current `center` placement also centers an opaque 760-pixel panel over a presenter. For a horizontal talking-head video with the speaker in the middle, this is the opposite of a center-safe layout.
-
-## Audience and source contract
-
-- Primary source: 1920x1080 horizontal talking-head footage.
-- Presenter: normally centered.
-- Existing burned-in captions remain authoritative and are never duplicated.
-- Source video and source audio stay continuous unless a future job explicitly authorizes a recut.
-- Graphics explain, structure, emphasize, or show evidence; they do not invent claims.
-
-## Layout contract
-
-### Protected regions
-
-- Presenter-safe center: x = 35% to 65% of frame width.
-- Left overlay lane: x = 6% to 32%.
-- Right overlay lane: x = 68% to 94%.
-- Burned-in subtitle band: bottom 18% of the frame.
-- Ordinary overlays must not place opaque panels or essential text in the presenter-safe center or subtitle band.
-
-### Allowed coverage
-
-- `speaker-left`: content occupies the left lane and leaves the center transparent.
-- `speaker-right`: content occupies the right lane and leaves the center transparent.
-- `full-screen`: content may cover the presenter for 1.0-2.2 seconds and is reserved for an opening hook, chapter transition, or closing payoff.
-- Evidence takeover may exceed 2.2 seconds only when the source requires readable evidence and the Gate A plan explicitly records the reason.
-
-### Rhythm
-
-- Side lanes alternate; the same side cannot appear twice consecutively unless one intervening beat is a full-screen scene or breathing beat.
-- One template appears at most three times and never on adjacent beats.
-- No more than two strong or full-screen scenes in succession.
-- The planner must not assign a graphic to every subtitle merely to fill time.
-
-## Visual families
-
-### 1. Impact Question
-
-- Use: opening question, contradiction, major turn.
-- Motion: question hit, counter-line slide, conclusion stamp.
-- Palette: Deep Ocean (`#0B1325`, `#FF6846`, `#55D7FF`).
-- Coverage: full-screen only at approved hook/turn boundaries.
-
-### 2. Gradient Keyword
-
-- Use: one dominant term or short conclusion.
-- Motion: gradient charge, word lift, underline sweep.
-- Palette variants: Violet Sunset and Cobalt Magenta.
-- Coverage: speaker-left or speaker-right; no opaque center panel.
-
-### 3. Signal Route
-
-- Use: workflow, connection, feedback loop, tool chain.
-- Motion: route trace, node activation, pulse travel.
-- Palette: Teal Signal (`#071B22`, `#55D7FF`, `#54F2D2`) or Coral Signal.
-- Coverage: two side clusters connected around, not through, the presenter-safe center.
-
-### 4. Editorial Stamp
-
-- Use: correction, judgment, locked conclusion.
-- Motion: editorial line reveal followed by a physical stamp.
-- Palette: Editorial Cream (`#F4F0E8`, `#18181B`, `#D92D20`, `#2563EB`).
-- Coverage: speaker-left/right for ordinary judgments; full-screen only for a chapter conclusion.
-
-### 5. Completion Rail
-
-- Use: execution checkpoints, QA, next action, CTA.
-- Motion: ordered rails fill and the final state stamps in.
-- Palette: Acid Action (`#111827`, `#C7F000`, `#FACC15`) or Navy Cyan.
-- Coverage: edge rails above the subtitle band; center remains transparent except for an approved closing payoff.
-
-### 6. Programmatic Doodle
-
-- Use: abstract mechanism, difficulty, tradeoff, speed, path, or before/after.
-- Motion: deterministic layered SVG pieces driven by frame progress.
-- Palette: warm paper, relaxed black line, one low-saturation semantic accent.
-- Coverage: side illustration or brief full-screen chapter illustration.
-- Boundary: a drawing is explanatory and is never presented as evidence.
-
-## Palette system
-
-Palette selection is semantic, not random. Each storyboard beat stores a palette identifier. The renderer resolves that identifier through one registry so Remotion and the optional adapter share the same colors.
-
-Required palettes:
-
-1. `deep-ocean`: navy, coral, ice blue.
-2. `violet-sunset`: violet, magenta, warm orange.
-3. `teal-signal`: dark teal, mint, bright blue.
-4. `editorial-cream`: cream, ink, Chinese red, cobalt.
-5. `acid-action`: charcoal, acid green, yellow.
-6. `paper-sketch`: warm paper, ink, muted semantic accent.
-
-For a real case longer than 30 seconds, validation requires at least four palette identifiers and five visual structures. Adjacent beats may not repeat the same `structure + palette` combination.
-
-## Real-case direction: Auto Editing 0
-
-The authorized local MP4 and SRT remain the source for the public real-call case. The source files remain private.
-
-The revised sequence must contain:
-
-- a three-stage full-screen impact hook;
-- alternating left/right gradient-keyword and editorial beats;
-- one route or process scene;
-- one programmatic doodle scene;
-- one before/after or attributed metric treatment;
-- one completion/CTA scene;
-- at least four color families in the actual rendered output.
-
-The public preview must be extracted from the newly rendered MP4, not from isolated component mocks. It must include at least six frames covering a full-screen hook, left speaker-safe scene, right speaker-safe scene, signal route, doodle, and completion/closing scene.
-
-## Public documentation
-
-The README and Skill must state that the workflow targets center-presenter talking-head footage and must explain the two side lanes, protected center, subtitle band, and limited full-screen exceptions.
-
-The plan-first contract remains mandatory:
-
-1. inspect MP4/SRT and create Gate A;
-2. present the proposed beat map, layout, palette, factual boundaries, and evidence needs;
-3. stop;
-4. continue only after explicit approval for that job.
-
-## Validation
-
-Automated checks must cover:
-
-- generated 16:9 talking-head beats use only left, right, or approved full-screen placement;
-- ordinary beats never resolve to an opaque centered panel;
-- real cases over 30 seconds use at least four palettes and five structures;
-- adjacent beats do not repeat a `structure + palette` pair;
-- no template appears more than three times;
-- no two ordinary side beats use the same side consecutively;
-- full-screen use is limited to allowed director roles;
-- caption mode remains burned-in without a generated caption track;
-- all new SVG paths are safe from default black fills;
-- actual output decodes with ffmpeg, matches source duration tolerance, has no sampled black frames, and supplies verified preview hashes;
-- public repository scan finds no private source media, credentials, cookies, tokens, local browser state, or private-tool wording.
-
-## Non-goals
-
-- Reproducing an entire authoring application.
-- Publishing private source footage or SRT content.
-- Adding payment, deployment, production data, or online account access.
-- Creating dozens of nearly identical skins.
-- Claiming drawings, counters, or generic graphics as proof.

package/skills/package-talking-head-video/docs/superpowers/specs/2026-08-08-semantic-template-system-v2-design.md

@@ -1,365 +0,0 @@
-# Semantic Template System V2 — Design Specification
-
-**Status:** Implemented and verified against the `auto-editing-0` real invocation
-**Target:** Public `ai-self-media-video-packaging-skill` repository
-**Primary use case:** 16:9, center-framed talking-head footage with burned-in or separately managed captions
-
-## 1. Problem statement
-
-The current public example demonstrates color variation, but it does not yet demonstrate enough structural or editorial variation. Several scenes reuse the same visual grammar with different colors, so the result reads like decorated text rather than designed video packaging.
-
-V2 must improve the result at the template-system level. It must select a structure because that structure matches the meaning of the spoken segment, not because a palette rotation is due.
-
-The public package remains self-contained: the repository must include every rule, component, token, test, and example artifact required to reproduce its advertised result.
-
-## 2. Goals
-
-V2 must:
-
-1. Provide ten genuinely different semantic structures instead of aliases around a few card layouts.
-2. Preserve the on-camera speaker when the person occupies the center of the frame.
-3. Use the left and right sides as designed editorial rails, not loose text dumping zones.
-4. Reserve full-screen graphics for moments where replacing or dimming the A-roll materially improves comprehension.
-5. Generate programmatic line illustrations when an abstract concept is clearer as an action or process.
-6. Plan first, require explicit approval, and only then render.
-7. Re-run the `auto-editing-0` case and publish screenshots taken from that real run.
-8. Keep Remotion as the default renderer and keep the HyperFrames adapter available as an optional compatible output path.
-
-## 3. Non-goals
-
-V2 does not include:
-
-- a graphical editing application;
-- automatic publishing, uploading, payments, or account operations;
-- unrestricted template cloning or one-to-one reproduction of third-party artwork;
-- fabricated metrics, testimonials, product capabilities, or source attributions;
-- bundled source footage, source subtitles, credentials, cookies, or machine-specific paths.
-
-## 4. Required workflow and approval gates
-
-Every invocation follows four explicit gates:
-
-### Gate A — Analysis and proposal
-
-The Skill inspects media metadata and subtitles, identifies semantic beats, proposes structures, flags factual risks, and writes a human-readable plan. It must not render the final result at this gate.
-
-### Gate B — Approved storyboard
-
-After explicit approval, the Skill creates the deterministic timeline, fills template contracts, prepares allowed evidence or illustrations, and produces storyboard evidence. It does not claim final delivery.
-
-### Gate C — Interactive or frame-level review
-
-The Skill builds the real composition, runs automated checks, and produces representative review frames. Defects return to Gate B or implementation as appropriate.
-
-### Gate D — Final render
-
-Only an explicit Gate D approval permits final rendering. The produced file must be inspected as an artifact, not inferred from a successful command.
-
-## 5. Frame and safety model
-
-The reference canvas is 1920×1080.
-
-### Center-presenter mode
-
-- Protected presenter lane: `x = 35%–65%`.
-- Left editorial lane: `x = 5%–32%`.
-- Right editorial lane: `x = 68%–95%`.
-- Bottom caption reserve: the lower `18%` of the frame unless the source is explicitly caption-free.
-- Critical text, evidence, diagrams, and counters must not enter the protected presenter or caption lanes.
-- No foreground surface, text, path, glow, or illustration may cross the protected lane.
-
-### Full-screen mode
-
-Full-screen treatment is permitted only for:
-
-- the opening hook;
-- a process that requires three or more connected stages;
-- before/after comparison;
-- evidence inspection;
-- a numeric conclusion;
-- the final summary or CTA.
-
-Full-screen scenes use an opaque edge-to-edge background. There is no translucent hybrid mode: if meaningful geometry enters the presenter lane, the presenter is intentionally replaced for that beat.
-
-### Side-card density
-
-- Side surfaces follow content height and never fill the lane from a fixed top to bottom.
-- Maximum card height is 64% of the frame.
-- One idea uses one compact rail; two rails require two genuinely complementary groups.
-- Representative review frames are sampled at 72% of the beat, then manually checked for face safety, empty space, clipping, contrast, semantic fit, and subtitle clearance.
-
-## 6. Semantic template library
-
-Each template is a distinct component with its own content schema, layout, and motion contract. Palette changes do not create a new structure.
-
-### 6.1 `editorial-dual-rail`
-
-**Use for:** a main argument with two complementary groups of supporting points.
-**Layout:** center presenter; framed left and right rails; top kicker; footer takeaway.
-**Required fields:** `kicker`, `headline`, `leftItems[1..3]`, `rightItems[1..3]`, `takeaway`.
-**Motion:** header settles first, rails enter from opposite sides, numbered items reveal in reading order, takeaway locks last.
-**Do not use for:** a sequential process or direct comparison.
-
-### 6.2 `thesis-and-proof`
-
-**Use for:** one strong claim that requires a concise reason, source, or observation.
-**Layout:** large thesis on one side; proof block with source label on the other; center presenter remains open.
-**Required fields:** `thesis`, `reason`, optional `sourceLabel`, optional `sourceDetail`.
-**Motion:** thesis hits, reason lifts, source stamp appears only after the proof is readable.
-**Do not use for:** unsupported promotional claims.
-
-### 6.3 `bidirectional-flow`
-
-**Use for:** input/output, tool/human, problem/solution, or two-way feedback relationships.
-**Layout:** left and right nodes connected around the presenter lane; return path is visually distinct.
-**Required fields:** `leftLabel`, `rightLabel`, `forwardAction`, optional `returnAction`, `result`.
-**Motion:** forward route draws first; return route draws only when the narration establishes a feedback loop.
-**Do not use for:** a one-direction timeline.
-
-### 6.4 `command-palette`
-
-**Use for:** executable actions, commands, shortcuts, or a small operational checklist.
-**Layout:** compact command surface on one side, output or status summary on the other.
-**Required fields:** `commandTitle`, `actions[2..5]`, `resultState`.
-**Motion:** focus bar moves through actions; result state stamps after the final action.
-**Do not use for:** general benefits or conceptual nouns.
-
-### 6.5 `four-stage-pipeline`
-
-**Use for:** exactly three or four ordered operational stages.
-**Layout:** full-screen or side-spanning connected stages with a visible current-stage indicator.
-**Required fields:** `title`, `stages[3..4]`, optional `output`.
-**Motion:** route draws once; stages activate in chronological order; completed stages remain legible.
-**Do not use for:** unordered lists.
-
-### 6.6 `before-after-scrub`
-
-**Use for:** exactly two comparable states with a stable comparison dimension.
-**Layout:** split field or scrub bar; labels remain anchored; the comparison criterion stays visible.
-**Required fields:** `before`, `after`, `criterion`, optional `delta`.
-**Motion:** reveal progresses from before to after; delta appears only if it is sourced or explicitly presented as an estimate.
-**Do not use for:** three alternatives or non-comparable examples.
-
-### 6.7 `evidence-panel`
-
-**Use for:** screenshots, documentation, research excerpts, or verifiable artifacts.
-**Layout:** evidence image or excerpt receives the largest area; interpretation and source are secondary.
-**Required fields:** `evidenceAsset`, `caption`, `sourceLabel`, `interpretation`.
-**Motion:** evidence enters first, crop or highlight follows, interpretation appears last.
-**Do not use for:** decorative stock imagery or invented proof.
-
-### 6.8 `metric-odometer`
-
-**Use for:** one to three real numeric values central to the spoken point.
-**Layout:** large number, unit, meaning label, and source/estimate status.
-**Required fields:** `metrics[1..3]`, each with `value`, `unit`, `label`, and `evidenceStatus`.
-**Motion:** numbers count only when interpolation is honest; otherwise use a direct hit or step change.
-**Do not use for:** numbers without a source or clearly stated owner estimate.
-
-### 6.9 `signal-route`
-
-**Use for:** a system path, dependency chain, handoff, or data route with three to five nodes.
-**Layout:** route avoids the presenter lane or uses full-screen dim mode; nodes have distinct roles.
-**Required fields:** `nodes[3..5]`, `routeLabel`, optional `failureNode`, `result`.
-**Motion:** line draw, node activation, optional failure pulse, result lock.
-**Do not use for:** a simple list disguised as a diagram.
-
-### 6.10 `semantic-doodle`
-
-**Use for:** an abstract idea best explained through a human action, transformation, or cause/effect metaphor.
-**Layout:** one clear subject, one action, and one outcome; avoid white-card default framing.
-**Required fields:** `subject`, `action`, `outcome`, `accent`, optional `annotation`.
-**Motion:** line path draws, local color blocks reveal, subject acts, result appears.
-**Do not use for:** generic decoration or a literal restatement of the subtitle.
-
-## 7. Visual system
-
-### Typography hierarchy
-
-- Kicker or category: compact, high-contrast, visually subordinate.
-- Display claim: the largest text in the frame, normally one sentence or two short lines.
-- Card heading: visually distinct from body copy.
-- Body copy: short clauses, not paragraphs.
-- Source or status label: always readable and never used as decoration.
-
-Default copy limits:
-
-- Display claim: maximum 24 Chinese characters or 80 Latin characters.
-- Card heading: maximum 14 Chinese characters or 48 Latin characters.
-- Body item: maximum 22 Chinese characters or 72 Latin characters.
-- Source label: maximum 28 Chinese characters or 90 Latin characters.
-
-### Color
-
-- Each scene uses a base, one dominant accent, and at most one support accent.
-- Adjacent scenes must not differ only by hue.
-- Palette choice follows semantic tone: operational, evidentiary, comparative, cautionary, or conclusive.
-- Text contrast must meet WCAG AA: 4.5:1 for normal text and 3:1 for large text.
-
-### Surfaces and depth
-
-- Side rails use controlled transparency, borders, editorial rules, and layered spacing.
-- Full-screen structures may use paper, console, blueprint, or evidence surfaces when semantically appropriate.
-- Shadows, glow, grain, and gradients are support devices, never substitutes for hierarchy.
-
-## 8. Motion system
-
-All default Remotion motion is frame-derived and deterministic.
-
-- Normal scene target: 3–5 seconds.
-- Hard scene limit: 6 seconds unless evidence readability requires a documented exception.
-- Entry or emphasis phase: normally 0.4–0.8 seconds.
-- No wall-clock timers or runtime-only CSS animation dependencies.
-- Every scene must render correctly when seeking directly to any frame.
-- Exiting elements may not leave important content partially clipped at scene boundaries.
-
-Illustration rules:
-
-- SVG paths explicitly set `fill="none"` where appropriate.
-- Line drawing uses deterministic dash offsets.
-- Color fills are local accents, not full-card recolors.
-- At least three visual layers must participate: line, subject/action, and outcome/annotation.
-
-## 9. Director selection rules
-
-The director pipeline is:
-
-1. classify the subtitle beat;
-2. identify eligible structures;
-3. reject structures whose content contracts cannot be satisfied;
-4. choose placement from speaker and caption safety metadata;
-5. score cadence diversity without overriding semantic fitness;
-6. write the Gate A proposal;
-7. materialize only after approval;
-8. validate and render.
-
-Selection triggers:
-
-- `step`, `stage`, `first/next/final` → `four-stage-pipeline` or `signal-route`;
-- `before/after`, `instead`, `from/to` → `before-after-scrub`;
-- `according to`, `documentation`, `screenshot`, `evidence` → `evidence-panel`;
-- a central numeric statement → `metric-odometer`;
-- an executable sequence → `command-palette`;
-- a claim plus support → `thesis-and-proof`;
-- complementary categories → `editorial-dual-rail`;
-- two-way handoff or feedback → `bidirectional-flow`;
-- abstract human action or transformation → `semantic-doodle`.
-
-Cadence rules:
-
-- The same structure may not appear more than twice consecutively.
-- A video longer than 30 seconds should normally use at least eight distinct structures when the semantics support them.
-- Left and right emphasis should alternate over time.
-- Full-screen structures should not appear consecutively unless they form one deliberate sequence.
-- Diversity checks may flag a weak plan, but they must not force a semantically incorrect template.
-
-## 10. Code architecture
-
-### Core planner
-
-The core package owns:
-
-- beat classification;
-- template eligibility and scoring;
-- content-schema validation;
-- center-presenter and caption-safe layout metadata;
-- gate manifests;
-- factual-risk labels.
-
-### Remotion renderer
-
-The Remotion package owns:
-
-- ten independent template components;
-- shared typography, tokens, safe-area helpers, evidence surfaces, and SVG primitives;
-- deterministic frame-based motion;
-- the default final renderer.
-
-Existing family aliases that map different names to the same layout must be removed or retained only as compatibility shims that resolve to a documented V2 structure.
-
-### HyperFrames adapter
-
-The adapter remains supported as an optional output path. It must preserve the same manifest, timing, safe areas, evidence status, and semantic structure identifiers. Its presence does not replace the Remotion implementation or its tests.
-
-## 11. Real-call example
-
-The `auto-editing-0` example will be regenerated with V2.
-
-Required outputs:
-
-- Gate A analysis and approved plan;
-- machine-readable timeline manifest;
-- final Remotion render generated from the real example invocation;
-- eight representative screenshots covering at least eight structures;
-- one contact sheet assembled from those exact frames;
-- artifact metadata and checksums;
-- a short factual note distinguishing source claims, owner-confirmed claims, and estimates.
-
-The existing preview set must be replaced rather than presented alongside the new evidence. README images must point only to the new verified screenshots.
-
-## 12. Test strategy
-
-Implementation follows test-driven development.
-
-### Unit and contract tests
-
-- every template has a distinct schema and renderer identity;
-- semantic triggers map to eligible templates;
-- invalid template/meaning combinations are rejected;
-- copy-length, evidence-status, presenter-safe, and caption-safe rules fail loudly;
-- same-structure and palette-only repetition is detected;
-- SVG illustration paths preserve explicit fill and stroke semantics.
-
-### Composition tests
-
-- all ten structures appear in the Remotion composition catalogue;
-- direct seek snapshots are stable;
-- center-presenter and subtitle reserves contain no critical overlay bounds;
-- contrast meets required thresholds;
-- no important content clips or overflows.
-
-### Real artifact tests
-
-- final render exists and has expected non-zero size;
-- `ffprobe` confirms intended resolution, frame rate, duration, video, and audio streams;
-- full-file decode completes without error;
-- black-frame and frozen-frame checks pass or produce reviewed exceptions;
-- eight representative screenshots are manually reviewed against their semantic contracts;
-- published screenshots are confirmed to come from the final artifact or the exact approved composition state.
-
-## 13. Public repository hygiene
-
-Before any push:
-
-- scan tracked files for `.env`, keys, tokens, cookies, private keys, credentials, absolute machine paths, source media, source subtitle files, temporary browser state, and runtime caches;
-- verify that all screenshots are intended for public release;
-- keep example input media excluded from Git;
-- include licenses and attribution for any reused open-source code or assets;
-- ensure README claims match verified capabilities;
-- retain the approved final business contact `nanaya093` only in the requested closing copy.
-
-## 14. Acceptance criteria
-
-V2 is accepted only when all of the following are true:
-
-1. Ten independent semantic structures exist and are documented.
-2. At least eight structures are visibly present in the regenerated real example.
-3. The example is not dominated by free-floating colored text or one repeated card layout.
-4. The center presenter and burned-in subtitle areas remain readable throughout manual review.
-5. Programmatic doodles demonstrate line drawing, local color, action, and result.
-6. Gate A is generated before implementation and final rendering requires explicit Gate D approval.
-7. Unit tests, typecheck, build, public-repository scan, composition checks, and artifact checks pass.
-8. README preview images are replaced with screenshots from the regenerated real invocation.
-9. No source media, source subtitles, credentials, machine-specific paths, or unintended internal material are tracked.
-10. The requested README closing copy, including business WeChat `nanaya093`, remains unchanged.
-
-## 15. Delivery sequence
-
-1. Approve this written specification.
-2. Produce a file-by-file implementation plan.
-3. Add failing contract tests for semantic structures and safety rules.
-4. Implement the core planner and ten Remotion templates.
-5. Update the HyperFrames adapter contract.
-6. Run the real `auto-editing-0` example through Gate A and request approval.
-7. After approval, produce Gate B/C evidence and request Gate D.
-8. Render, inspect, replace previews, scan the repository, and publish only after final verification.

package/skills/package-talking-head-video/examples/auto-editing-0/GATE_A.md

@@ -1,56 +0,0 @@
-# 自动剪辑 0｜V2 真实调用 Gate A
-
-状态：`GATE_A_APPROVED__HISTORICAL_RECORD`
-
-本文件记录 Gate A 当时完成的真实输入探测、逐条 SRT 解析、语义导演方案和事实边界审计。该方案随后已获批准，并依次完成 Gate B、Gate C 与 Gate D；最终结果见 [GATE_D.md](GATE_D.md) 与 [QA_REPORT.md](QA_REPORT.md)。
-
-## 输入基线
-
-- 视频：49.273 秒，1920×1080，30fps，HEVC + AAC。
-- 视频 SHA-256：`3CAEFB03337D6E9A1357F7437507B73868E269A63DFCA0DA5BC9B458B8F6E696`
-- SRT：21 条，源片已烧录字幕，本次使用 `burned-in` 模式，不重复生成字幕。
-- SRT SHA-256：`93968AF7002BECE6B4E4C1A4841171BAB9D43F38EE25D82D2CFC06E7F31ECC40`
-- 原始视频与 SRT 只位于忽略目录，不进入公开仓库。
-
-## 方案结果
-
-- 22 个可读视觉状态；一条短桥接字幕与下一句合并，一条长指标句和一条长预告句被拆分。
-- 8 种语义结构：`thesis-and-proof`、`editorial-dual-rail`、`command-palette`、`metric-odometer`、`semantic-doodle`、`before-after-scrub`、`four-stage-pipeline`、`signal-route`。
-- 6 套配色：深海蓝、紫橙渐变、青绿信号、米白编辑、荧光行动、纸张简笔画。
-- 最长单一视觉状态 3.200 秒。
-- 左侧 6 段、右侧 6 段、全屏 10 段；左右侧信息连续序列 0 次同侧重复。
-- 画面中央 35%–65% 为人物安全区；侧栏只使用 5%–32% 与 68%–95%；底部 18% 保留给源片烧录字幕。
-
-## 全屏结构时间码
-
-| 时间 | 结构 | 目的 |
-|---|---|---|
-| 00:00.000–00:02.666 | 观点 + 解释 | 三秒内建立“一键智能剪辑”钩子 |
-| 00:04.867–00:07.000 | 指标大字 | 将“几万流量池”明确标为个人案例 |
-| 00:14.766–00:15.933 | 语义简笔画 | 创作者跨过反复踩坑，沉淀可复用 Skill |
-| 00:15.933–00:18.033 | 指标大字 | 展示“迭代几个月”的本人经历 |
-| 00:18.100–00:20.133 | 前后对照 | 人工找素材 vs 原视频进入工作流 |
-| 00:20.133–00:22.033 | 四阶段流程 | 找素材、上字幕、卡点、加特效 |
-| 00:22.133–00:24.700 | 前后对照 | 2–3 小时仍难完成 vs 交给工作流 |
-| 00:24.700–00:27.866 | 信号路径 | 放入原视频 → 工作流处理 → 约 5 分钟完成 |
-| 00:29.400–00:31.800 | 指标大字 | Plus 周额度 3%，仅按本人当次口径表达 |
-| 00:32.066–00:34.200 | 指标大字 | “一周几十条”明确标为估算 |
-
-## 左右侧信息时间码
-
-- 左侧：00:02.733–00:04.867、00:08.266–00:11.433、00:27.866–00:29.266、00:36.566–00:39.533、00:41.933–00:44.216、00:46.600–00:48.100。
-- 右侧：00:07.133–00:08.266、00:11.500–00:14.700、00:34.366–00:36.533、00:39.600–00:41.833、00:44.216–00:46.500、00:48.100–00:49.266。
-
-## 事实与证据边界
-
-- “几万流量池”“迭代几个月”“约 5 分钟”“Plus 周额度 3%”只作为讲述者个人案例或当次使用口径，不升级为普遍效果承诺。
-- “一周几十条”标为估算，不作为稳定产能或性能承诺。
-- 口播中的“爆款概率拉满”不在包装文案中放大；视觉文案改为“剪辑提升内容完成度 / 爆款仍取决于选题、文案与平台反馈”。
-- 当前输入没有提供可公开、可追溯的批准证据截图，因此方案中不使用 `evidence-panel`，也不虚构来源。
-- 资料领取、关注和下期预告只复述源片口播，不新增联系方式、收费信息或发布承诺。
-
-## Gate 结论
-
-- Gate A：通过并获批准。
-- Gate B/C：随后已完成；八张代表帧均从最终 MP4 抽取并公开。
-- Gate D：随后已完成；最终成片通过完整解码与黑帧扫描，但完整 MP4 不随公开仓库分发。

package/skills/package-talking-head-video/examples/auto-editing-0/GATE_D.md

@@ -1,8 +0,0 @@
-# Gate D — final render accepted
-
-- Gate A: approved; source probe, full SRT parse, semantic plan and fact boundaries completed before implementation.
-- Gate B: approved; 22 seek-safe visual states implemented with 8 semantic structures and 6 palettes.
-- Gate C: approved; representative frames reviewed for presenter safety, caption safety, semantic clarity and layout diversity.
-- Gate D: approved; final Remotion artifact rendered and verified.
-
-Final artifact facts are recorded in [QA_REPORT.md](QA_REPORT.md) and [preview-manifest.json](preview-manifest.json). The full MP4 and source media are intentionally not distributed in this public repository.

package/skills/package-talking-head-video/examples/auto-editing-0/QA_REPORT.md

@@ -1,34 +0,0 @@
-# QA report
-
-## Decisive artifact
-
-- Renderer: Remotion.
-- Output SHA-256: `60BE56C417FE63EF17F18874A4EE74318D001CD40D7A40BA3312488900B85443`.
-- Size: 32,921,365 bytes.
-- Duration: 49.322667 seconds.
-- Video: H.264, 1920×1080, 30fps.
-- Audio: AAC, 48kHz, stereo.
-
-## Checks performed
-
-- Remotion rendered 1,479/1,479 frames.
-- Full FFmpeg decode completed without an error.
-- Black-frame scan returned no black interval.
-- The continuous source picture and audio remain present for the complete timeline.
-- The source already contained subtitles; no duplicate caption layer was generated.
-- Eight representative frames and an eight-frame contact sheet were extracted from the actual final MP4, not from a component preview.
-- The approved plan uses 6 palettes, 8 actual structures, 22 visual beats, 6 left-lane placements, 6 right-lane placements, and 10 short full-screen placements; the longest state is 3.2 seconds.
-- Human inspection confirmed that presenter-safe frames leave the complete center 35%–65% lane unobstructed and keep the bottom burned-in captions readable.
-- Human inspection confirmed that every structure entering the center lane uses a 100% opaque edge-to-edge full-screen surface; no translucent hybrid covers the presenter.
-- Human inspection confirmed that side cards fit their content, remain below the 64% height cap, and do not contain a large empty lower half.
-- Verbal metrics such as “几万” remain literal instead of being coerced to `0`.
-- The semantic doodle visibly depicts a person pushing a boulder uphill rather than a generic symbol.
-- Signal-route nodes and result labels use an opaque full-canvas explanation state instead of partially covering the presenter.
-- All eight representative screenshots were sampled at 72% of their beat, after the layout had reached its stable readable state.
-- The line illustration uses deterministic SVG paths with explicit `fill="none"`.
-- Claims in the source narration remain attributed to the speaker; the renderer adds no invented percentage, testimonial, logo, or performance claim.
-- Original MP4 and SRT are absent from the public repository candidate set.
-
-## Reproduction note
-
-The public repository includes the source hashes and the exact repository-relative invocation, but not the copyrighted or person-identifying source files. A user can reproduce the workflow with their own MP4 and SRT using the same command shape.

package/skills/package-talking-head-video/references/director-rules.md

@@ -1,29 +0,0 @@
-# Director and evidence rules
-
-## Semantic direction
-
-- Establish one readable tension, question, or result in the first three seconds.
-- Change structure when the reasoning changes: claim, proof, loop, command, sequence, comparison, evidence, metric, route, or metaphor.
-- Do not animate every noun. One visual state should carry one meaningful relationship.
-- Avoid more than two consecutive instances of the same structure. Alternate ordinary left/right lanes.
-- A video longer than 30 seconds needs at least eight valid structures; diversity must come from the speech, not random assignment.
-
-## Center-presenter protection
-
-- Protect center x=35%–65% and bottom 18% by default.
-- Protect faces, products, source UI, nameplates, and burned-in captions after inspecting actual frames.
-- A full-screen beat must preserve a presenter window or intentionally dim the source; it must not accidentally cover the person with dense copy.
-
-## Facts and evidence
-
-- A spoken number remains a speaker claim unless independently verified.
-- Mark numbers as `sourced`, `owner-confirmed`, or `estimate`; never silently promote one status to another.
-- `evidence-panel` requires an approved local screenshot/document, a source label, and a restrained interpretation.
-- If the source asset is absent, downgrade to `thesis-and-proof`; do not fabricate a proof card.
-- A doodle explains action or transformation. It never functions as evidence.
-
-## Source continuity
-
-- Preserve the source edit and audio by default.
-- A browser-compatible H.264/AAC proxy may be generated without cuts when the source codec cannot be decoded in the renderer.
-- If captions are already burned in, do not add another caption track.