PAPER ALGORITHM · 纸上算法
From one sentence to a full production line — making AI clear for everyone.

English · 简体中文

---

🧭 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.

Paper is the stage, ink is the text, vermilion is the accent, the seal is the signature. Three skills, each guarding one page — together they form a single moving page of algorithm notes: from "what should I make this week" to delivered videos, automatic email notices, and a four-platform publishing schedule.

⏱ 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.

Check your setup — run python -X utf8 scripts/check_skills.py (should output All checks passed) and python scripts/doctor.py (lists missing tools, cross-platform) before producing your first video. See CONTRIBUTING.md for how to contribute, CHANGELOG.md for recent changes, and the Code of Conduct.

---

Pipeline Runner (Optional)

After installation, use the pipeline runner instead of manually executing steps. The runner runs 6 step scripts in order, checks preconditions at each step, and stops on failure:

# List all steps
python -m pipeline.runner --list

# Preview execution plan
python -m pipeline.runner --dry-run

# Run steps (step 3 confirmation gate requires --confirm to pass)
python -m pipeline.runner --project-dir <project-dir> --steps 1,2

# Continue after confirmation
python -m pipeline.runner --project-dir <project-dir> --steps 3,4,5,6

Six steps: 01 optimize → 02 check_prose → 03 confirm (hard gate) → 04 tts → 05 render → 06 verify (hard gate).

---

📑 Contents

• What Is This
• 30-Second Quick Start
• Pipeline Runner
• Three Pages · Skill Overview
• Showcase
• The Voiceover Workflow
• Paper Algorithm · Design System
• Voice Presets
• Themes
• For AI Agents
• The Aliveness Gate
• Installation and Usage
• Repository Layout
• Use and Remix
• License

---

🧩 Three Pages · Skill Overview

🎬 Voiceover Router tryworld-koubo One sentence in, auto-routed through topics, script, polish, and video production. Delivers: full pipeline + email notice $tryworld-koubo

📜 Paper Algorithm Video tryworld-paper Script → branded horizontal video, a page of moving algorithm notes. Delivers: video · covers · titles · captions · multi-voice $tryworld-paper

🎯 AI Topic Selection tryworld-topics Hundreds of daily AI headlines, distilled into 3-8 topics worth making — fresh, catchy, not repeated. Delivers: topic list (angle · source · priority) $tryworld-topics

---

🎬 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"]) --> P["koubo Router"]
    P -- "Mode A" --> A["paper Video"]
    P -- "Mode B" --> S["topics Picks"]
    S --> A
    A --> O["Deliverables"]
    O --> M["Email Notice"]

Loading

---

🎨 Paper Algorithm · Design System

The visual contract behind every TryWorld video — scientific manuscript + Chinese print tradition, a locked contract with no shortcuts.

Paper #F4EFE4 Background

Ink #1C1916 Text & lines

Vermillion #C0452F The only accent

Ink Blue #2E5E8C Notes & charts

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.

---

🎨 Themes

The default look is the Paper Algorithm theme (TryWorld's brand), but all brand values — palette, fonts, seal, platform name, sign-off, cover style, default voice — live in one theme file: skills/tryworld-paper/themes/paper-algorithm.json.

Want your own brand on this pipeline? Copy the theme JSON, change the values following tryworld-paper/references/theme-guide.md, and ship videos under your name. The pipeline contract is not configurable — confirmation gates, the aliveness-gate check, burned-in captions, and quality checks all stay — only the skin changes. Content theme (content-default.json) defines topic domain, audience, length range, and target platforms — modify this JSON to change domain/brand, SKILL.md references it automatically. Writing methodology is in tryworld-koubo/references/writing-methodology.md; set "writing_rules": "custom" in the content theme to replace with your own methodology (but the check_prose.py hard gate cannot be disabled). The repo includes test-brand.json as a complete rebranding example.

// themes/your-brand.json (excerpt)
{
  "brand": { "platform_name": "YourChannel", "seal_text": "原创认证", "sign_off": "..." },
  "palette": { "background": "#F7F7F2", "text": "#111111", "accent": "#0F62FE", "secondary": "#5A6B7B" }
}

Community themes are welcome — PR to skills/tryworld-paper/themes/ with a 720p sample (CC BY-SA 4.0 applies).

---

🤖 For AI Agents

If you are an AI coding agent (Claude Code / Codex / Cursor …) whose user installed or cloned this repo, this section is your operating manual. The human talks; you drive the pipeline.

1. Skills are executable manuals

Each folder under skills/ is a skill whose SKILL.md is a complete, executable instruction set:

Skill	Role	Read it when
skills/tryworld-koubo/SKILL.md	Single entry / router	Any voiceover request (make an episode, pick topics, optimize a script)
skills/tryworld-paper/SKILL.md	Script → video production	Once a script is confirmed for production
skills/tryworld-topics/SKILL.md	AIHOT topic selection	Mode B topic picking

After installation into ~/.agents/skills, hosts trigger these by the description frontmatter; when the user names one ($tryworld-koubo), open its SKILL.md and follow it literally.

2. The execution sequence for one full episode

1. Route — read tryworld-koubo/SKILL.md, decide Mode A (user gave a script → optimize & produce) or Mode B (topic request → select → write → produce).
2. Environment check — node --version (≥22) · python --version (≥3.10) · python -c "import edge_tts" · ffmpeg -version · npx hyperframes doctor. Install or warn on whatever is missing before proceeding.
3. Mode B only — run python tryworld-topics/scripts/fetch_aihot.py (cross-platform) or the curl fallback in its references/api.md, apply references/selection-rules.md, produce 3–8 candidate topics.
4. Gate 1 (STOP) — present the topic list and wait for the user's pick. Do not proceed unasked.
5. Write — follow the writing rules in koubo's SKILL.md (2,500–2,800 chars standard, sources for every datapoint, aliveness rules).
6. Polish & gate — hand over to tryworld-paper/SKILL.md: optimize, then run scripts/check_prose.py until zero hard violations.
7. Gate 2 (STOP) — show the full polished script + what/why of edits + element mapping + data sources. Wait for explicit confirmation. Never render without it.
8. Produce — read the theme file (skills/tryworld-paper/themes/paper-algorithm.json) for all brand values → segment → scripts/tts_yunxi.py (voiceover + sentences.json timeline) → build HyperFrames compositions per references/style-system.md → npx hyperframes lint / validate / inspect --strict all pass → render (draft first, then high) → covers (independent design, never video frames) → titles per references/titles.md → everything into outputs/ + 发布计划.txt.
9. Notify — run python tryworld-koubo/scripts/notify_delivery.py --project-dir <dir>; missing credentials/node → skip gracefully, never block delivery.
10. Verify before delivery (hard gate) — run python -X utf8 tryworld-paper/scripts/verify_output.py --dir <outputs>; every item must PASS (audio track present, durations aligned, both cover sizes, titles, publish plan). A FAIL means fix and re-run — never deliver an unverified outputs folder.

Projects live in their own folder (default convention: E:\Codex口播视频\<slug>\ with work/ and outputs/).

3. Hard rules (violating any = task failed)

• Both gates require the user's explicit confirmation. Never self-advance past a gate.
• check_prose.py hard violations must be zero before the polished script may even be shown.
• Brand values (palette, seal, sign-off, fonts, default voice) come only from the theme file — never hardcode them from memory.
• Before rendering, lint/validate/inspect --strict must pass with zero errors and warnings.
• Script data points need sources; unverifiable ones are marked 待核实, never invented.
• Run verify_output.py on outputs/ before handing anything to the user; a silent video or a missing cover is a failed delivery.

4. Common failures

Symptom	Fix
npx hyperframes missing	First npx call auto-installs it; needs Node ≥22
edge-tts 403 / rate limit	The tts script auto-retries once on failure; the AIHOT fetch script retries 429/5xx with backoff — retry or switch network
Email notice fails	Credentials missing → it skips by design; delivery is not blocked
Fonts wrong in render	Check style-system.md font embed table; non-native fonts need bundled woff2

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/skills convention (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 that tryworld-koubo routes 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

In-repo check scripts

• python -X utf8 scripts/check_skills.py: compiles skill Python scripts and runs the aliveness gate
• python -m pytest tests/ -v: run 23 unit tests (check_prose, verify_output, tts_yunxi, check_skills) over examples/ scripts and titles.
• python scripts/doctor.py: checks Python, Node, FFmpeg/ffprobe, edge-tts, HyperFrames, and optional email credentials. Exits 1 when a required item is missing.

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:

• Cross-platform — data-fetching (tryworld-topics) and email-notice (tryworld-koubo) scripts are Python (.ps1 fallback versions retained); tryworld-topics documents a curl fallback in its references/.
• Python 3.10+ — pip install edge-tts for the voiceover pipeline (tryworld-paper/scripts/tts_yunxi.py, default Yunxi + voice presets); tryworld-paper/scripts/check_prose.py needs 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_CODE credentials (via the $qq-email skill); skipped gracefully when absent.

Each skill folder carries its own SKILL.md with the full workflow and where to change environment defaults.

---

Docker One-Shot Environment

Skip all dependency installs with Docker:

docker build -t paper-algorithm .
docker run --rm paper-algorithm

Container includes Node.js 22, FFmpeg, Python 3.12, edge-tts, and mutagen.

---

🗂 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)
├── scripts/                      # In-repo checks (static gates + doctor)
├── 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
    │   ├── themes/              # Brand theme files (paper-algorithm.json)
    └── 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.

---

_{TryWorld · Making AI clear for everyone}