📚 Academic Writing Agents: AI Specialists for Research Paper Workflows

From draft to submission — 11 specialized AI agents that handle translation, polishing, de-AI-ification, experiment analysis, diagram design, and adversarial peer review for academic papers.

中文版 README

---

📖 Why This Project?

When you're debugging the same polishing prompt for the third time, the lab next door has already finished three papers with ready-made templates.

In academia, prompt engineering is becoming a hidden resource — top research groups have polished template libraries, while most people are still figuring things out from scratch. We don't want this inequality to continue.

awesome-ai-research-writing took the first step: collecting battle-tested prompts from researchers at MSRA, ByteDance Seed, Shanghai AI Lab, Peking University, USTC, and SJTU, then open-sourcing them for everyone.

This project takes the next step: transforming those scattered prompts into structured AI agents using the Agency Agents format — so they can be loaded into Claude Code, Cursor, Copilot, and 6 other tools as persistent specialists with identity, workflow, rules, and success metrics.

Stop spending time on prompt debugging. Save your energy for actual research.

---

🚀 What Is This?

A collection of 11 specialized AI agent personalities for academic writing. Each agent is:

• 🎯 Specialized: One agent, one job — translation, polishing, de-AI, review, etc.
• 🧠 Personality-Driven: Not generic "helpful assistant" prompts, but opinionated specialists with strong voices and distinct philosophies
• 📋 Prompt-Embedded: Original Chinese prompts from real CS/ML researchers, embedded verbatim and ready to use
• 🔬 Battle-Tested: Refined through real paper preparation workflows at top research institutions
• ✅ Measurable: Concrete success criteria with numeric thresholds (e.g., false positive rate < 5%, detection rate > 90%)

---

⚡ Quick Start

Option 1: Use with Claude Code (Recommended)

# Copy agents to your Claude Code directory
cp agents/*.md ~/.claude/agents/

# Then activate in Claude Code:
# "Activate Academic Translator mode and translate this paragraph"
# "Use the Peer Review Simulator agent to review my paper"
# "Switch to AI Writing Decontaminator and clean up this section"

Option 2: Use with Cursor / Copilot / Other Tools

These agents follow the Agency Agents format and are compatible with its conversion pipeline:

# Clone the agency-agents repo and copy these agents into it
cp agents/*.md /path/to/agency-agents/academic/

# Then convert for your tool
cd /path/to/agency-agents
./scripts/convert.sh --tool cursor    # For Cursor (.mdc rules)
./scripts/convert.sh --tool aider     # For Aider (CONVENTIONS.md)
./scripts/convert.sh --tool windsurf  # For Windsurf (.windsurfrules)

# Install interactively
./scripts/install.sh

Supported tools: Claude Code, GitHub Copilot, Cursor, Aider, Windsurf, Gemini CLI, OpenCode, OpenClaw, Qwen Code.

Option 3: Use the Prompts Directly

Each agent file contains the original Chinese prompts inside code blocks in the Technical Deliverables section. You can extract and use them directly in ChatGPT, Gemini, or any other LLM — no agent framework required.

For example, open academic-translator.md, find the ## 📋 Your Technical Deliverables section, and copy the prompt inside the code fence.

---

🎓 The Agent Roster

#	Agent	Specialty	When to Use
🌉	Academic Translator	Bidirectional Chinese-English academic translation	Translating Chinese drafts to publication-ready English LaTeX, or English papers to Chinese for comprehension
🏗️	Chinese Academic Restructurer	Transforms fragmented notes into coherent paragraphs	Rewriting scattered bullet points or oral-style drafts into formal Chinese academic text for Word
✒️	English Academic Polisher	Deep English LaTeX polishing for top conferences	Achieving zero-error publication quality for NeurIPS, ICLR, ICML submissions
🪶	Chinese Academic Polisher	Conservative Chinese editing for Word manuscripts	Minimal-intervention polishing that respects the author's original voice
⚖️	Text Length Optimizer	Precision 5-15 word compression or expansion	Fine-tuning paragraph length without losing information or adding filler
🔍	AI Writing Decontaminator	Detect and neutralize AI-generated writing patterns	Making AI-assisted text indistinguishable from human-written prose (English + Chinese)
🛡️	Logic Auditor	Final red-line logic verification	Pre-submission check for fatal contradictions, terminology swaps, ambiguous grammar
🎨	Paper Architecture Illustrator	Academic method diagram prompts (DeepMind aesthetic)	Generating prompts for architecture figures with flat vector style and pastel palette
📊	Experiment Data Specialist	Chart type recommendation + analysis paragraph writing	Choosing from 19 academic chart types and writing LaTeX analysis with \paragraph{} format
🏷️	Academic Caption Writer	Figure and table caption generation	Converting Chinese descriptions to publication-ready English captions with correct casing
⚔️	Peer Review Simulator	Adversarial pre-submission review simulation	Simulating harsh top-conference critique to find fatal flaws before real reviewers do

---

🔄 Recommended Workflow

A typical paper preparation pipeline using these agents:

                        ┌─────────────────────┐
                        │   Raw Chinese Draft  │
                        └──────────┬──────────┘
                                   │
                    ┌──────────────┼──────────────┐
                    ▼              ▼               ▼
            ┌──────────┐  ┌──────────────┐  ┌──────────┐
            │ 🏗️ Re-   │  │ 🪶 Chinese   │  │ 🌉 Trans-│
            │ structurer│  │ Polisher     │  │ lator    │
            └──────┬───┘  └──────┬───────┘  └────┬─────┘
                   │             │               │
                   └─────────────┼───────────────┘
                                 ▼
                        ┌────────────────┐
                        │  English LaTeX │
                        └───────┬────────┘
                                │
              ┌─────────────────┼─────────────────┐
              ▼                 ▼                  ▼
       ┌──────────┐     ┌──────────┐       ┌──────────┐
       │ ✒️ English│     │ ⚖️ Length │       │ 🔍 De-AI │
       │ Polisher  │     │ Optimizer│       │          │
       └─────┬────┘     └─────┬────┘       └─────┬────┘
             │                │                   │
             └────────────────┼───────────────────┘
                              ▼
                     ┌────────────────┐
                     │  Polished LaTeX│
                     └───────┬────────┘
                             │
           ┌─────────────────┼──────────────────┐
           ▼                 ▼                   ▼
    ┌──────────┐     ┌──────────────┐    ┌──────────┐
    │ 🛡️ Logic │     │ 🎨 Arch Illus│    │ 📊 Experi│
    │ Auditor  │     │ + 🏷️ Caption │    │ ment     │
    └─────┬────┘     └──────┬───────┘    └─────┬────┘
          │                 │                   │
          └─────────────────┼───────────────────┘
                            ▼
                   ┌─────────────────┐
                   │  Final Manuscript│
                   └────────┬────────┘
                            │
                            ▼
                   ┌─────────────────┐
                   │ ⚔️ Peer Review  │
                   │ (Pre-submission) │
                   └─────────────────┘

---

💡 Usage Examples

Example 1: Translate a Chinese Draft for ICML

Activate Academic Translator mode.
Translate the following Chinese draft into English LaTeX for ICML submission:

我们提出了一种新的注意力机制，通过动态调整查询和键的交互方式来提升模型在长序列上的效率。
与标准的多头注意力相比，我们的方法在保持相同精度的同时将计算复杂度从 O(n²) 降低到 O(n log n)。

Output: 3-part response — LaTeX code, Chinese back-translation, modification log.

Example 2: Remove AI-Generated Patterns

Switch to AI Writing Decontaminator mode (English).
Clean this paragraph:

This paper delves into the intricacies of leveraging large language models
to foster more nuanced understanding. Our approach unveils a novel paradigm
that fundamentally transcends existing methodologies.

Output: Rewritten text with "delves", "leveraging", "foster", "unveils", "paradigm", "transcends" replaced with natural academic alternatives.

Example 3: Pre-submission Peer Review

Activate Peer Review Simulator mode.
Review my uploaded PDF. Target venue: NeurIPS 2026.

Output: Part 1 (Review Report with Summary, Strengths, Critical Weaknesses, Rating) + Part 2 (Strategic Advice with actionable revision guidance). All in Chinese.

---

🎭 Agent Personality Highlights

"I am not your friend. I am the reviewer who will reject your paper if you do not fix this."

— Peer Review Simulator ⚔️

"The best edit is the one you do not make."

— Chinese Academic Polisher 🪶

"I am not here to improve your paper. I am here to save it from rejection."

— Logic Auditor 🛡️

"You read text the way a forgery expert reads handwriting."

— AI Writing Decontaminator 🔍

"You are not a polisher — you are an architect."

— Chinese Academic Restructurer 🏗️

"Data does not speak for itself — I give it a voice that is precise, honest, and compelling."

— Experiment Data Specialist 📊

---

🗺️ Prompt-to-Agent Mapping

How the 16 original prompts from awesome-ai-research-writing map to these 11 agents:

Original Prompt	Agent	Strategy
中转英 + 英转中	Academic Translator	Merged — bidirectional, two modes
中转中	Chinese Academic Restructurer	Standalone — rebuilds, not polishes
表达润色（英文）	English Academic Polisher	Standalone — aggressive deep rewriting
表达润色（中文）	Chinese Academic Polisher	Standalone — conservative minimal editing
缩写 + 扩写	Text Length Optimizer	Merged — opposite operations, shared logic
去AI味（英文）+ 去AI味（中文）	AI Writing Decontaminator	Merged — same goal, two languages
逻辑检查	Logic Auditor	Standalone — narrow surgical scope
论文架构图	Paper Architecture Illustrator	Standalone — visual domain
实验绘图推荐 + 实验分析	Experiment Data Specialist	Merged — chart selection + analysis writing
生成图标题 + 生成表标题	Academic Caption Writer	Merged — near-identical structure
Reviewer 视角审视	Peer Review Simulator	Standalone — strongest personality

---

📋 Agent Format

Each agent follows the Agency Agents standard with 9 sections:

---
name: Agent Name
description: One-line specialty
color: "#hexcode"
emoji: <emoji>
vibe: One-line personality hook
---

# Agent Name
## 🧠 Your Identity & Memory       ← Who you are
## 🎯 Your Core Mission             ← What you do
## 🚨 Critical Rules You Must Follow ← Hard constraints
## 📋 Your Technical Deliverables    ← Embedded prompts (Chinese, verbatim)
## 🔄 Your Workflow Process          ← Step-by-step
## 💭 Your Communication Style       ← How you communicate
## 🔄 Learning & Memory             ← What you track across sessions
## 🎯 Your Success Metrics           ← Measurable outcomes with numeric thresholds
## 🚀 Advanced Capabilities          ← Extended features

The original Chinese prompts from awesome-ai-research-writing are embedded verbatim inside code fences in the Technical Deliverables section of each agent.

---

📊 Stats

• 🎓 11 Specialized Agents for academic writing
• 📝 16 Battle-tested Prompts from top CS/ML research groups (MSRA, Seed, Shanghai AI Lab, PKU, USTC, SJTU)
• 🌐 Bilingual — English agent shell + Chinese embedded prompts
• 🔧 9 Tool Integrations — Claude Code, Cursor, Copilot, Aider, Windsurf, and more
• 📄 ~1,900 lines of personality, process, and domain expertise

---

🤝 Contributing

We welcome contributions! You can:

1. Improve existing agents — sharpen rules, add examples, refine metrics
2. Add new agents — writing-related specialists we haven't covered (e.g., related work writer, rebuttal drafter, abstract generator)
3. Report issues — prompts that produce suboptimal results in specific scenarios
4. Share workflows — how you chain these agents in your paper preparation pipeline

Please follow the Agency Agents format when contributing new agents.

---

🙏 Acknowledgments

This project builds on two open-source foundations:

• awesome-ai-research-writing — The upstream prompt source, collected from researchers at MSRA, ByteDance Seed, Shanghai AI Lab, Peking University, USTC, SJTU, and other institutions. All Chinese prompts in this project are preserved verbatim from the original.

• The Agency (agency-agents) by @msitarzewski — The agent format, design philosophy, and multi-tool integration architecture. Our agents follow the Agency standard and are fully compatible with its conversion and installation pipeline.

---

📜 License

MIT License — Use freely, commercially or personally. Attribution appreciated but not required.

---

📚 Academic Writing Agents: Your AI Research Team 📚

⭐ Star this repo · 🍴 Fork it · 🐛 Report an issue