🧭 What Is This
试界TryWorld official AI-agent skill set — a complete Chinese voiceover-video production line, from one sentence to a delivered video. Topic selection, scriptwriting, polish, rendering, publish plan, and email notice are all packaged into three Skills:
- 🎬
tryworld-koubo— the single entry. One sentence in, auto-routed through topics → script → polish → video. - 🎯
tryworld-topics— AIHOT-based topic selection: hundreds of daily AI headlines distilled into 3–8 fresh picks. - 📜
tryworld-paper— the "Paper Algorithm" video pipeline: script → branded 16:9 video with covers, titles, and captions.
Install them into ~/.agents/skills, open a session in any agent host that reads that directory (e.g. Agent SDK / ZCode), and say one sentence in Chinese. The skills pick up the whole chain from there — topic list, scripts, rendered video, covers, titles, even the automatic email notice.
⏱ 30-Second Quick Start
1 · Install — clone the repo, then copy the skill folders into your skills directory:
git clone https://github.com/TryWorld2026/paper-algorithm.git
cd paper-algorithm
mkdir -p ~/.agents/skills
cp -r skills/* ~/.agents/skills/
On Windows PowerShell:
git clone https://github.com/TryWorld2026/paper-algorithm.git
cd paper-algorithm
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.agents\skills"
Copy-Item -Path .\skills\* -Destination "$env:USERPROFILE\.agents\skills" -Recurse
2 · Say it — open a session with your agent and just talk:
帮我做一期口播
That's it. The request is recognized and routed through the whole pipeline automatically — no commands to memorize. You can also name a skill directly ($tryworld-koubo, $tryworld-topics, $tryworld-paper); the $-name is the skill's call sign in agent sessions.
3 · Receive — the pipeline ends with the rendered video, horizontal/vertical covers, platform titles, captions, a four-platform publish plan, and an automatic email notice.
📑 Contents
- What Is This
- 30-Second Quick Start
- Three Pages · Skill Overview
- Showcase
- The Voiceover Workflow
- Paper Algorithm · Design System
- Voice Presets
- The Aliveness Gate
- Installation and Usage
- Repository Layout
- Use and Remix
- License
🧩 Three Pages · Skill Overview
🎬 Showcase
Real outputs from the pipeline — everything below was produced end-to-end by these skills, with no manual editing. Videos are 720p previews (originals are ~30–45 MB, 1080p); each folder also carries the voiceover script, platform titles, and the publish plan.
| Example | Topic | Length | Video preview | In the folder |
|---|---|---|---|---|
| cordis-paper | DeepSeek 论文精读:把拆插件讲成数学 | 8.8 min | preview_720p.mp4 · 12 MB | covers · titles · script · publish plan |
| cursor-spacex | SpaceX 买下 Cursor | 6.6 min | preview_720p.mp4 · 8 MB | covers · titles · script · publish plan |
🎬 The Voiceover Workflow
Remember one entry: $tryworld-koubo. Hand it a script (Mode A) or ask for topics (Mode B).
flowchart LR
U(["One sentence from you"]) --> P["🎬 tryworld-koubo<br/>Single entry, auto route"]
P -- "Mode A, script given" --> A["📜 tryworld-paper<br/>Optimize, confirm, produce"]
P -- "Mode B, pick a topic" --> S["🎯 tryworld-topics<br/>AIHOT news to topic list"]
S -- "you choose" --> A
A --> O["✅ Deliverables<br/>Video, covers, titles, schedule"]
O -. automatic .-> M["📧 Email notice<br/>Deliverables and publish times"]
🎨 Paper Algorithm · Design System
The visual contract behind every TryWorld video — scientific manuscript + Chinese print tradition, a locked contract with no shortcuts.
| Dimension | Convention |
|---|---|
| Type | Noto Serif SC (headings) · ZCOOL XiaoWei (notes) · monospace (data) |
| Motion | ink drop · brush stroke · seal stamp — three signature moves throughout |
| Authenticity | vermilion「试界原创」seal, always visible top-right — the only watermark |
🎙 Voice Presets
The voiceover defaults to Yunxi (云希, TryWorld's brand voice), with seven more presets built in — one flag to switch:
| Preset | Voice | Character | Audition |
|---|---|---|---|
yunxi (default) |
云希 | male, sunny | ▶ listen |
xiaoxiao |
晓晓 | female, warm | ▶ listen |
xiaoyi |
晓伊 | female, lively | ▶ listen |
yunjian |
云健 | male, passionate | ▶ listen |
yunyang |
云扬 | male, news anchor | ▶ listen |
yunxia |
云夏 | male, youthful | ▶ listen |
xiaobei |
小贝 | female, Northeastern dialect | ▶ listen |
xiaoni |
小妮 | female, Shaanxi dialect | ▶ listen |
All samples above are synthesized with TryWorld's fixed sign-off line, so you hear each voice exactly as it sounds in a real video.
python tryworld-paper/scripts/tts_yunxi.py script.txt --out work/audio --voice xiaoxiao
Full edge-tts voice ids are accepted too. By default the pipeline never asks about voices — say you want a different one (a female voice, a news-anchor feel) and it synthesizes short auditions from your own script for you to pick.
🔒 The Aliveness Gate
Before delivery, every polished script and platform title passes a machine gate — the ban targets rhetorical moves, not literal strings. Repeating the same move with different words still counts.
- Flip-flop rhetoric: sets up a misunderstanding the reader never had, then overturns it for dramatic lift. Known guises include 不是……而是……, 并非……而是……, 表面……实际……, 看似……实则……, 你以为……其实……, 回头才发现, 说到底, 答案恰恰相反 — state judgments directly, judgment first, evidence after.
- Triple+ parallel structure: three or more identical constructions; keep at most two.
- Lyric metaphor: no concrete verbs bolted onto abstract nouns ("time keeps the details" type); unaffected when writing about concrete things.
- Nominalization: "实现了效率的提升" → say how much faster, how many people saved.
- Punctuation tiers: all dashes banned; colons only to introduce direct speech.
- Jargon tiers: absolute bans + context-sensitive words, maintained by the checker.
tryworld-paper/scripts/check_prose.py (from human-writing v1.1.0, MIT) runs these checks automatically and adds statistical signals — sentence-length variance, conjunction density, model-favorite lyric words, 「」-quote density. Zero hard violations required before the user-confirmation gate; failing means no delivery.
🛠 Installation and Usage
Install
Each skill folder is a self-contained Skill. Copy it into your local skills directory:
# Install all skills
Copy-Item -Path .\skills\* -Destination "$env:USERPROFILE\.agents\skills" -Recurse
# Or install a single skill
Copy-Item -Path .\skills\tryworld-paper -Destination "$env:USERPROFILE\.agents\skills" -Recurse
Hosts that support the
~/.agents/skillsconvention (e.g. Agent SDK / ZCode) pick these up automatically; for other hosts, point your host's skills directory at this path (see your host's docs). Note thattryworld-kouboroutes to its sibling skills, so for the full pipeline install all three — see Environment for external dependencies.
How to trigger
A skill fires when the request matches its trigger words — no commands to memorize:
- Say it in plain words: 「帮我做一期口播」「这周做什么口播」「把这篇稿子做成视频」…
- Name the skill explicitly:
$tryworld-koubo/$tryworld-topics/$tryworld-paper - Paste a script and ask to produce it — the router detects the input and goes straight into the video pipeline.
What you get
| You say | Route | You get |
|---|---|---|
| 帮我做一期口播 / 帮我选题 | $tryworld-koubo (auto) |
routed pipeline · publish plan · email notice |
| 这周做什么口播 | via koubo → tryworld-topics | 3–8 topic picks (angle · source · priority) |
| 把这篇稿子做成视频 | via koubo → tryworld-paper | 16:9 video · covers · titles · captions |
Environment
Run npx hyperframes doctor to check the environment at once. These skills are built for TryWorld's own production workflow, so a few things are assumed — all adjustable in the skill files:
- Windows PowerShell — data-fetching (
tryworld-topics) and email-notice (tryworld-koubo) scripts are PowerShell (scripts/*.ps1);tryworld-topicsdocuments a curl fallback for non-Windows in itsreferences/. - Python 3.10+ —
pip install edge-ttsfor the voiceover pipeline (tryworld-paper/scripts/tts_yunxi.py, default Yunxi + voice presets);tryworld-paper/scripts/check_prose.pyneeds no third-party packages. - Node.js >= 22 + HyperFrames — rendering (
npx hyperframes render/lint/validate/inspect). - FFmpeg (with ffprobe, on PATH) — audio processing.
- External skills —
$hyperframes(rendering, required) and$qq-email(email notice, optional) are not in this repo; install them separately. - Working directory — finished projects land in
E:\Codex口播视频by convention; change it in the skill files to your own. - Email notice — requires
QQ_EMAIL_ACCOUNT/QQ_EMAIL_AUTH_CODEcredentials (via the$qq-emailskill); skipped gracefully when absent.
Each skill folder carries its own SKILL.md with the full workflow and where to change environment defaults.
🗂 Repository Layout
paper-algorithm/
├── assets/
│ ├── hero.svg # Brand banner
│ └── license-badge.svg # License badge
├── README.md # Index (English)
├── README.zh-CN.md # Index (简体中文)
├── LICENSE # CC BY-SA 4.0
├── examples/ # Real pipeline outputs (showcase)
│ ├── cordis-paper/ # Sample: 720p preview · covers · titles · script
│ └── cursor-spacex/ # Sample: 720p preview · covers · titles · script
└── skills/
├── tryworld-koubo/ # Voiceover router (routing + email notice)
├── tryworld-paper/ # Paper Algorithm video production
└── tryworld-topics/ # AIHOT topic selection
🤝 Use and Remix
- Fork, don't copy. Use the fork button — it keeps the link back to this repository and its commit history.
- Keep the upstream. Every derivative repository must carry a visible attribution to this project (
TryWorld2026/paper-algorithm) and the license notice. - Share alike. Remixes must be released under the same license (CC BY-SA 4.0). A repository that doesn't carry this license statement is not licensed for derivative work.
- Commercial use is welcome — as long as the attribution and the license travel with it.
📄 License
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license: share and adapt freely — including commercially — as long as you give credit, and any remix is released under the same license. Full terms: LICENSE.
tryworld-paper/scripts/check_prose.py is derived from human-writing v1.1.0 and remains under the MIT License (see skills/tryworld-paper/LICENSE-MIT); everything else in this repository is CC BY-SA 4.0.
No comments yet
Be the first to share your take.