qiaomu-cut
Verified把一句话需求转成可复现、可验收视频工程的乔木智能剪辑导演。Use when the user asks to create, plan, edit, remix, explain, narrate, subtitle, animate, composite, or render a video—including one-line requests such as “制作一个科普视频:介绍 LLM 中的 RL”. Builds English-learning movie mixes, person profiles, explainers, code-drawn motion graphics (UI morph loops, showreels, kinetic-type promos, line-art/pixel/Three.js animation, lyric MVs), product launch videos, cinematic shorts, social clips, course/PPT videos, AI image/video/TTS/music-assisted stories, stock-footage stories, or multi-source montage. Routes across 33台词, ClipSeek/open media, local files, ListenHub, image generation, HTML/SVG/Canvas motion graphics with beat-grid storyboards and determinism probes, Manim, ffmpeg-full, captions, sound design, frame-0 video covers with YouTube thumbnails, and release checks.
agents/interface.yaml
View on GitHubname: qiaomu-cut
version: 0.16.0
mode: governed
summary: 一句话生成专业视频的乔木智能剪辑导演
owner:
name: 向阳乔木
x: https://x.com/vista8
github: https://github.com/joeseesun/
intents:
- id: create_video
description: 从自然语言 brief 规划、采集、生成、剪辑和验证视频
examples:
- 给我做一个 60 秒挖掘机英语启蒙视频
- 搜几段电影里骂人的台词,混剪成英语学习视频
- 介绍乔布斯,做一个电影感人物短片
- 把这段口播剪成小红书风格短视频
- 制作一个科普视频:介绍 LLM 中的 RL
- 做一个 15 秒的动态图形 showreel
- 用代码做一个卡 120 BPM 的 UI 形态变换循环动效
- id: source_media
description: 搜索、生成或整理视频、照片、插画、影视台词、旁白、音乐、本地文件和 AI 素材
- id: plan_video
description: 生成 QiaoCut IR、分镜、镜头语言、转场和渲染计划
- id: verify_video
description: 检查输出视频编码、时长、分辨率、音频和来源记录
inputs:
brief:
type: string
required: true
description: 用户的一句话视频需求
workflow:
type: enum
required: false
values:
- english-mix
- stock-story
- person-profile
- explainer
- explainer-social
- cinematic-short
- product-launch
- social-short
- talking-head
- data-story
- motion-design
- hybrid-studio
assets:
type: array
required: false
description: 本地文件、URL、素材候选或已下载素材
output:
type: object
required: false
properties:
aspect: "9:16 | 16:9 | 1:1 | 21:9"
durationSeconds: number
platform: "douyin | xiaohongshu | tiktok | youtube | course | custom"
profile: "preview | standard | final"
outputs:
- qiaocut-ir.json
- assets-manifest.json
- license-report.md
- quality-report.json
- final.preview.mp4
- final.standard.mp4
- final.mp4
creative_defaults:
narration:
provider_priority: ListenHub
preferred_speaker_name: 向阳乔木 v1.1
speaker_resolution: 默认由 qcut listenhub narration 执行,--voice-name 默认即“向阳乔木 v1.1”,调用时无需指定;只接受唯一、已授权的“向阳乔木 v1.1”完全同名 speaker ID,并自动记录 speaker/text/catalog provenance
format: mp3
failure_policy: 找不到、歧义或未授权时不悄悄换音色;回退用户已提供的 file narration 或请求选择
billing_policy: 有 estimate 先 estimate;当前 OpenAPI TTS 无已验证 estimate,必须说明费用未知并为本次调用取得 --yes
localization: 二进制输出必须项目相对、no-clobber,生成后立即 ingest;URL 结果立即 fetch
scene_motion:
focus_spotlight: primaryFocus 渲染为真实聚光:非焦点元素压暗降饱和,焦点光晕+微脉冲;focus-sequence 不画矩形光晕圈,强调由步骤激活态承担
visible_change_semantics: state=焦点 from→to 文本变形+高亮环;process=序列节点全程逐步激活+连接符生长+落位 settle 微振;两者都随场景时长铺排,无死时间
beats: "qcut explainer beats --apply 按旁白静音分句点自动生成;spec 场景可选 [{at: 秒, action: activate|swap|pulse, target?: 步骤序号}] 手填覆盖"
review: scene-review 首/中/尾三帧;state|process 场景自动首尾帧差检测,YAVG<2.0 判 visibleChangeVerified=false
narration_gate: spec 旁白锁 timing 但 timeline.narration.engine=none 时 preview/final 报错;final 抽听成片中段 RMS<-40dB 报错
music_default: explainer 知识片默认 music:false,程序化合成垫不作默认交付
generated_images:
style_inputs:
- content_theme
- audience
- era
- emotion
- medium
consistency: 先建立整片 visual bible,后续图像继承统一色板、镜头/构图、光线、纹理、字体/图形语言和人物约束
mismatch_policy: 不硬编码单一风格;提示词与主题或 visual bible 不匹配时拒绝生成并先重写/重规划
commands:
setup:
run: node scripts/qcut.js setup --json
description: 幂等安装缺失依赖(ffmpeg-full、中文字幕字体、playwright+Chromium、Manim、ListenHub CLI);doctor 报缺时 agent 直接运行后继续,不跳过、不降级
motion_init:
run: node scripts/qcut.js motion init ./promo --style product-promo --audio assets/song.mp3 --json
description: 生成六段式 motion-brief.json、节拍网格分镜骨架、pure-time 起步场景和 motion kit
motion_check:
run: node scripts/qcut.js motion check ./promo --json
description: 先分镜后代码硬门:镜头写完、切点落在节拍网格、Banned≥3、镜头连续、确定性 lint
motion_probe:
run: node scripts/qcut.js motion probe ./promo --json
description: 切点/中点/events 截图、正序倒序冷启动三次截图字节一致、循环接缝 PSNR、contact sheet
motion_render:
run: node scripts/qcut.js motion render ./promo --profile final --json
description: draft 半分辨率无模糊;final 需 probe 通过,3 子帧动态模糊、峰值对齐音效、−14 LUFS 迭代母带
audio_beats:
run: node scripts/qcut.js audio beats assets/song.mp3 --output reports/beat-grid.json --json
description: 无依赖节拍网格:BPM、样本级校准的拍点、强拍、小节能量与 drop
audio_peak:
run: node scripts/qcut.js audio peak assets/sfx/click.wav --json
description: 测音效峰值偏移,用于把峰值而不是文件起点对齐到画面事件
doctor:
run: node scripts/qcut.js doctor --json
description: 检查 ffmpeg-full、33tc、ClipSeek、本地工具能力
clipseek:
run: node scripts/qcut.js clipseek "挖掘机" --type video --json
description: 搜索 ClipSeek 免费素材候选
plan:
run: node scripts/qcut.js plan "做一个挖掘机英语启蒙视频" --json
description: 生成 QiaoCut IR 草案
verify:
run: node scripts/qcut.js verify /path/to/video.mp4 --json
description: 验证外部生成、移动后或单独收到的视频;qcut render 已内置对应档位校验,不重复执行
render_preview:
run: node scripts/qcut.js render /path/to/project --profile preview --json
description: Skill 默认第一步;960 长边/24 fps 上限、basic 校验、单遍响度、无 contact sheet
render_standard:
run: node scripts/qcut.js render /path/to/project --profile standard --json
description: 日常交付;1280 长边/30 fps 上限、响度/峰值/静音校验和最多 8 帧 contact sheet
render_final:
run: node scripts/qcut.js render /path/to/project --profile final --json
description: 正式终稿;timeline 原始输出、两遍响度、full 校验和 contact sheet;省略 profile 仍默认 final
33tc:
run: node scripts/qcut.js 33tc search "台词" --json
description: 委托独立安装且获授权的 33tc CLI;pick/cut 可能耗积分,必须显式确认;远程下载启用有界重连,已有 task ID 时用 download 恢复而不重新付费创建
english_mix_init:
run: node scripts/qcut.js english-mix init ./daily-english --phrases "That makes sense|I see what you mean" --clips-per-phrase 4 --json
description: 初始化影视英语混剪目录和目标数量明确的 source-selection.json
english_mix_audit:
run: node scripts/qcut.js english-mix audit ./daily-english --require-media --strict-boundaries --json
description: 在付费切片前校验多 cue 完整句上下文,下载后再检查音视频流、SHA-256 重复和字幕证据
english_mix_pace:
run: node scripts/qcut.js english-mix pace ./daily-english --crossfade-ms 200 --apply --json
description: 不覆盖原素材,先按字幕边界裁掉残句,再生成带 60ms 安静入场、500ms 静音尾帧和 40/160ms 音频淡入淡出的派生素材,验证边界低于 -60 dB,并约定 timeline 使用 200ms fade/audio acrossfade
english_mix_review:
run: node scripts/qcut.js english-mix review ./daily-english --frames 8 --json
description: 从预览自动生成最多 8 帧素材总览和最长英文字幕手机帧
scaffold:
run: node scripts/qcut.js scaffold ./qiaocut-project --brief "做一个 60 秒挖掘机英语启蒙视频" --json
description: 创建视频项目目录、IR、素材 manifest 和授权报告
ass:
run: node scripts/qcut.js ass captions.json --output subtitles.ass
description: 把 caption JSON 转成 ASS 字幕
html_scene:
run: node scripts/qcut.js html-scene qiaocut-ir.json --scene s01 --output scenes/s01.html
description: 把 IR scene 转成可捕获的 HTML 视觉场景
scene_render:
run: node scripts/qcut.js scene render ./project scenes/concept.html --engine html --output assets/scenes/concept.mp4 --duration 4 --width 1080 --height 1920 --fps 30
description: 把 HTML/SVG/Manim 场景渲染成可进入 timeline 的项目内 MP4;浏览器帧通过 image2pipe 直接进入 ffmpeg,不创建 PNG 帧目录
scene_batch:
run: node scripts/qcut.js scene batch ./project scenes-batch.json --json
description: 同构 HTML/SVG 卡片共用一次 Chrome 会话批量渲染,减少重复启动与漏产物
explainer_autopilot:
run: node scripts/qcut.js explainer init ./LLM-RL --topic "LLM 中的 RL" --json
description: 一句话科普入口;创建七段结构、无左色条视觉模板、author/review/final 门和增量场景计划
explainer_preview:
run: node scripts/qcut.js explainer preview ./LLM-RL --json
description: 按依赖顺序只渲染过期场景,执行 author/scenes gate,生成快速预览及每场景首尾双帧认知审查表
explainer_timing:
run: node scripts/qcut.js explainer timing ./LLM-RL --audio assets/generated/narration.mp3 --apply --json
description: 测量真实旁白时长并生成 narration-timing.json;优先用静音边界形成 phrase timing,证据不足时显式降级为 duration-level,不冒充逐词对齐
explainer_final:
run: node scripts/qcut.js explainer final ./LLM-RL --json
description: 只有 autopilot-review 全部通过后执行 final/full 并要求 releaseReady
license:
run: node scripts/qcut.js license assets-manifest.json --output license-report.md
description: 生成素材授权报告
bootstrap_macos:
run: scripts/bootstrap_macos.sh --check
description: 检查或安装 ffmpeg-full
listenhub_doctor:
run: node scripts/qcut.js listenhub doctor --json
description: 只读检查 ListenHub npm 包/CLI 协议、OpenAPI key 与 OAuth/internal 状态、图片/视频/TTS/音乐/内容解析/Voice 能力和 Coli ASR;不显示凭据
listenhub_generate:
run: node scripts/qcut.js listenhub openapi video create --prompt "shot description" --model doubao-seedance-2-pro --no-wait --json --qcut-project /path/to/project --yes
description: 使用 LISTENHUB_API_KEY 时委托 OpenAPI video 生成候选素材;先 estimate,远端创建必须当前明确确认,本地文件上传另需 --allow-upload
listenhub_poll:
run: node scripts/qcut.js listenhub openapi video get TASK_ID_FROM_SUBMIT --json --qcut-project /path/to/project
description: 用创建结果中的 task ID 只读轮询,并把结果自动捕获到项目私有 .qiaocut/jobs/listenhub/ 下
listenhub_narration:
run: node scripts/qcut.js listenhub narration --text-file scripts/narration.txt --qcut-project /path/to/project --yes --json
description: 默认中文讲解闭环:不传 --voice-name 即默认唯一精确匹配“向阳乔木 v1.1”,默认 MP3,执行签名校验、自动 ingest,并返回带 provenance 的 timelineNarration;费用未知时先确认本次 --yes,失败后不自动二次付费
listenhub_asr:
run: node scripts/qcut.js listenhub asr assets/interview.wav --model sensevoice --json --qcut-project /path/to/project
description: 使用 Coli 做本地全文 ASR;不把它描述为逐词字幕对齐
fetch_generated:
run: node scripts/qcut.js fetch /path/to/project --result .qiaocut/jobs/listenhub/poll-result.json --field result.imageUrl --kind image --provider listenhub --visual-bible-id vb-ID --prompt "actual prompt" --seed PROVIDER_SEED --json
description: 从 poll capture 安全下载生成物并写任务及 visual-bible/prompt/seed provenance,不把临时 URL 放进 timeline
ingest_generated:
run: node scripts/qcut.js ingest /path/to/project /path/to/generated.mp4 --kind video --provider listenhub --json
description: 导入已在本机的生成素材,SHA-256 去重、no-clobber、不改原文件
bootstrap_listenhub:
run: scripts/bootstrap_listenhub.sh --check
description: 严格检查或精确安装 @marswave/listenhub-cli@0.0.15(CLI 协议 0.1.0)与 @marswave/coli@0.0.20;ListenHub provider 需要 Node 20+
trust_boundary:
reads:
- local media files explicitly provided by the user
- local 33TaiCi login state through an independently installed 33tc adapter; qcut redacts structured credential fields/URLs and the external adapter must not emit unlabelled raw credentials
- public ClipSeek/Pexels/Pixabay-style source pages
writes:
- project output directory chosen by user or agent
- generated reports and manifests
network:
- public media discovery and cited information sources
- 33TaiCi service calls made by the external adapter when the user invokes qcut 33tc
- ListenHub/MarsWave remote services only after per-request confirmation; local reference uploads require a separate allow-upload gate
avoids:
- deleting original media
- printing secrets, cookies, tokens, or private account data
- claiming provider license without source-page evidence
- redistributing 33TaiCi private protocols as part of this public skill
- bundling, uploading, or redistributing fonts discovered on the user's machine
- accepting API keys as command arguments or writing them to jobs, manifests, logs, reports, or Git
- putting temporary signed provider URLs in public manifests or timelines
- writing ListenHub captures outside the project-private .qiaocut/jobs/listenhub/ directory
- silently substituting another voice when the exact 向阳乔木 v1.1 speaker is absent, ambiguous, or unauthorized
- using a colored vertical stripe on the left edge of a card, quote, callout, takeaway, or block as a hierarchy shortcut
- hard-coding one generated-image style regardless of the brief or generating prompts that conflict with the project visual bible
- executing nested vendored MarsWave skills, auto-upgrades, agent-memory persistence, or cola-avatar-pack deletion instructions
performance_workflow:
default_cli_profile: final
skill_sequence:
- derive a project visual bible from theme, audience, era, emotion, and medium before generating imagery
- render preview for content, caption, pacing, and composition iteration
- render standard only for internal or speed-sensitive delivery
- render final once content is locked and release evidence is required
cache:
path: .qiaocut/cache
enabled_by_default: true
reusable_artifacts:
- shot segments
- macOS narration cues
- captioned picture master
- private local caption-font copy
bypass: --no-cache
validation:
preview: basic stream, duration, dimensions, frame-rate, pixel-format, and non-empty-file checks
standard: preview checks plus final loudness, true peak, and silence boundaries
final: standard checks plus full black-frame scan
release_ready: only final + full + verification pass + verified caption font when captions exist
font_policy: Prefer project fontsDir; otherwise reuse an installed local Noto Sans CJK SC only inside the private project cache and never distribute it.
measured_baseline:
fixture: 60-second DIG bilingual demo on the development Mac
legacy_final_seconds: 71.6
preview_cold_seconds: 27.7
preview_warm_seconds: 3.3-4.1
standard_cold_seconds_tts_warm: 27.1
standard_warm_seconds: 4.1
final_cold_seconds_tts_warm: 65.6
final_warm_seconds: 8.5
no_duplicate_verify: Do not invoke qcut verify immediately after a successful qcut render of the same file.
release_gate:
required:
- validate_skill.py
- trigger_eval.py
- qcut doctor
- README review
- secret scan before GitHub publish
- final profile with full validation and releaseReady true for release video fixtures
missing_evidence:
- end-to-end render fixtures for every workflow
- provider-backed download adapters beyond ClipSeek discovery
- human-reviewed public release
- an official public 33TaiCi API or redistribution permission for a bundled zero-install adapter
- paid ListenHub provider generation tests
- provider data-retention, training-use, regional-processing, and downstream-deletion evidence
- clean-host ListenHub/Coli install verification outside the development Mac
interface:
display_name: "Qiaomu Cut"
short_description: "一句话生成可复现、可验收的视频工程"
default_prompt: "Use $qiaomu-cut to turn my one-line brief into a reproducible, verified video project: plan, source or generate media, edit, caption, mix and render with release checks."
compatibility:
canonical_format: "agent-skills"
adapter_targets: ["openai", "claude", "generic", "agent-skills-compatible"]
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |