From illusion to practical academic research workflow: I used OpenClaw Skills to build a paper tracking system

0. Origin: A lesson from almost crossing the line

In March 2026, I asked AI to write my paper for me. The process was smooth - the paper is clearly structured, the figures are nice, and the data looks very reasonable.

Then I discovered: **Those data were generated by AI, not from real experiments. **

92% success rate, scalability analysis of 500 drones - it seems impeccable, but it is all “reasonable deduction”. If you submit it like this, it is academic fraud.

This experience made me think about one thing completely: **The biggest danger in LLM is not not being able to answer the question correctly, but answering confidently but getting the answer wrong. In an academic setting, this confidence can be fatal. **

So I redesigned my literature research workflow. There is only one core principle:

**Do not generate any unverifiable content. All documents must be manually searched and truly exist; all data must be well-documented. **

This is the original design intention of the two Skills paper-research and paper-verifier.

1. The core of the problem: Why is AI’s direct literature search unreliable?

Google Scholar has strict anti-crawler mechanisms. Selenium/Playwright automated operations are prone to IP blocking, triggering verification codes, and unstable results. More importantly: The papers searched by AI may not exist at all - LLM is good at hallucinatory answers that “sound reasonable”. In literature reviews, this is the most dangerous minefield.

Common false literature routines:

Make up the title and author of the paper
Fabricating non-existent journal conferences
Exaggerating citation numbers or impact factors
Call ordinary journals the top journals in the area

Therefore, I chose a method that looks “stupid” but is absolutely reliable: Human-machine collaboration, people-centered.

2. Overall design: four-step closed-loop workflow

手动搜索文献
     ↓
交互式收集元数据
     ↓
LLM 辅助生成综述
     ↓
Zotero 统一管理
     ↓
paper-verifier 交叉验证（可选）

Core logic: People are responsible for “finding real things”, and LLM is responsible for “organizing them efficiently”. LLM never generates content, only organizes and presents real information entered by humans.

3. paper-research Skill: complete literature research workflow

3.1 Quick initialization

# 创建研究工作区
./scripts/setup_manual_search.sh ./my_research "UAV path planning"

This will generate three files in the specified directory:

my_research/
├── 搜索指南.md         # 告诉你去哪搜、怎么搜
├── 文献记录模板.json   # 论文元数据存储格式
└── 收集文献.sh         # 交互式收集脚本

3.2 Search Guide: Manual Search Guidance with Strategy`Search Guide.md` is not just “search Google Scholar” - it contains:

Recommended Database:

Google Scholar (comprehensive, most complete)
IEEE Xplore (engineering authority)
Web of Science (Top Journal Index)
ACM Digital Library (computer)

Search Strategy:

# 示例：2023-2025 年 UAV 路径规划相关一区论文
"UAV path planning" AND "low altitude" AND year:2023..2025
site:ieeexplore.ieee.org "urban air mobility"

Filtering criteria:

Sort by number of citations, giving priority to highly cited papers
Verify whether the journal is in District 1 of the Chinese Academy of Sciences (query at fenqubiao.com)
Record DOI or accessible link

Reference List for Area 1 of the Chinese Academy of Sciences (built into Skill):

Type	Representative Journal
IEEE Transactions	TRO, TITS, TAE, TCST
Robot Top Issues	Automatica, JFR, RAS
Aviation Summit	ICRA, IROS, AIAA SciTech

3.3 Interactive literature collection

There are two ways, and interactive scripting is recommended:

cd my_research
./收集文献.sh

The script will gradually prompt for information about each paper:

===== 文献收集 =====
论文标题: Multi-Agent Path Planning for UAV Swarms
作者: Zhang S, Li M, Wang W
年份: 2024
期刊/会议: IEEE Transactions on Robotics
DOI: 10.1109/TRO.2024.3391285
引用数: 45
关键词: UAV, path planning, multi-agent, reinforcement learning
一句话总结: 用集中式训练+分布式执行的框架解决
            无人机集群路径冲突问题...
继续添加下一篇? (y/n):

If it is too slow, you can also directly edit the documentation record template.json:

[
  {
    "title": "Multi-Agent Path Planning for UAV Swarms",
    "authors": ["Zhang S", "Li M", "Wang W"],
    "year": 2024,
    "venue": "IEEE Transactions on Robotics",
    "doi": "10.1109/TRO.2024.3391285",
    "citations": 45,
    "keywords": ["UAV", "path planning", "multi-agent"],
    "summary": "集中式训练+分布式执行框架解决无人机集群路径冲突"
  }
]

3.4 Generating Markdown from JSON Literature Review

After the collection is completed, generate a structured review with one click:

python3 scripts/paper_collection.py \
    --input 文献记录模板.json \
    --output-md 文献综述.md

The output format is roughly as follows:

## 1. 研究背景

### 1.1 无人机集群路径规划

无人机集群路径规划是城市低空空域管理的核心问题...
Zhang et al. (2024) 提出的集中式训练+分布式执行框架...

### 1.2 多智能体强化学习

MARL 是解决分布式协同决策的主流方法...

## 2. 方法论分类

### 2.1 基于优化的方法
...

### 2.2 基于学习的方法
...

## 3. 关键文献汇总

| 论文 | 年份 | venue | 方法 | 贡献 |
|------|------|-------|------|------|
| Zhang et al. | 2024 | TRO | MARL | 提出 MGAT-AC 架构 |
...

The role of LLM is perfectly reflected here - it is not to find papers for you, but to organize the real information collected manually into a structured review text**.

3.5 Zotero synchronization: establishing a personal literature database

Collected literature can be imported into Zotero with one click:

export ZOTERO_LIBRARY_ID="你的图书馆ID"
export ZOTERO_API_KEY="你的API Key"

python3 scripts/zotero_manager.py \
    --library-id $ZOTERO_LIBRARY_ID \
    --api-key $ZOTERO_API_KEY \
    batch-add --file 收集的文献.json

In this way, all manually collected real documents will be synchronized to Zotero to form a reusable personal document library. Every time you do new research, you can add incrementally to the existing libraries.

4. paper-verifier Skill: authenticity cross-validationAfter collecting the literature, it still needs to be verified. This is the second Skill - Thesis Authenticity Verification Tool.

4.1 Why is verification required?

Manual searches also have the possibility of errors:

DOI filled in incorrectly
The author’s name is spelled incorrectly
The journal name is written as an abbreviation instead of the full name
The number of citations was recorded incorrectly (by dozens of times)
I thought it was Zone 1 but it turned out to be just CCF-B.

Therefore, before finalizing the review, use paper-verifier to do a cross-check.

4.2 DOI + Metadata Verification

python3 scripts/verify_papers.py \
    --input papers.json \
    --output verification_report.md

papers.json format:

{
  "title": "Multi-Agent Path Planning for UAV Swarms",
  "authors": "Zhang S, Li M",
  "year": 2024,
  "venue": "IEEE Transactions on Robotics",
  "doi": "10.1109/TRO.2024.3391285",
  "citations": 45
}

The script will be checked through two APIs:

Crossref API — Get official metadata:

Does the title match exactly?
Is the list of authors correct?
Are the years consistent?
What is the full name of the journal?

Semantic Scholar API — Secondary Verification:

Whether the paper actually exists
What is the actual number of citations?
Obtain the abstract of the paper to check the content

4.3 Verification report format

## 核查报告

### ✓ Zhang et al. (2024) - TRO
- DOI: 10.1109/TRO.2024.3391285 → 有效
- 期刊: IEEE Transactions on Robotics → 中科院一区 ✓
- 引用数: 声称45 → 实际52（Semantic Scholar）
  ⚠️ 引用数有出入，差7次
- 作者: Zhang S, Li M, Wang W → 核对通过 ✓

### ✗ Li et al. (2023) - ICRA
- DOI: 10.1109/ICRA.2023.1001234 → 有效
- ⚠️ 论文标题不匹配：实际为 "Single-Agent ..."
  → 请核实是否填错了论文

In this way, any manual input errors can be caught promptly.

5. File structure of two Skills

~/.openclaw/workspace/skills/
├── paper-research/          # 文献研究工作流
│   ├── SKILL.md            # 使用说明
│   ├── references/
│   │   └── zotero_setup.md # Zotero 配置指南
│   └── scripts/
│       ├── setup_manual_search.sh    # 工作区初始化
│       ├── paper_collection.py       # 文献收集+综述生成
│       ├── zotero_manager.py         # Zotero API 同步
│       ├── search_scholar.py         # Scholar 搜索（需API）
│       ├── search_serpapi.py         # SerpAPI 搜索
│       └── generate_summary.py       # 综述文本生成
│
└── paper-verifier/         # 真实性核查
    ├── SKILL.md
    └── scripts/
        └── verify_papers.py # 核心核查脚本

6. Linkage with RAG knowledge base

The collected real documents can not only generate a review, but also enter the LLMRAG knowledge base for subsequent Q&A and writing assistance.

The entire linkage process:

手动搜索 → 收集到 JSON → 生成 Markdown 综述
    ↓
导入 Zotero（长期管理）
    ↓
添加到 LLMRA G知识库（向量检索）
    ↓
后续论文写作时，RAG 问答检索真实文献

In this way, literature research is not a one-time task, but a knowledge asset that can be accumulated, retrieved, and reused.

7. Summary of design philosophy

The design of this workflow answers a fundamental question: **Where are the boundaries of LLM in academic research? **

LLM is good at	LLM is not good at (people must come)
Collate and organize information	Determine whether the document actually exists
Generate structured text	Verify DOI / citation count accuracy
Find knowledge gaps and connections	Search authoritative databases
Polishing and rewriting	Deciding which papers deserve inclusion in the review

This is also in line with the lesson learned from my previous paper incident: Don’t let LLM produce anything that cannot be traced back to the source. Every data and every document must be traceable.

8. Future expansion directions

Access SerpAPI to implement semi-automatic Google Scholar search (reduce manual operations)
Two-way synchronization with Zotero API to automatically pull metadata of read papers
Build a domain-specific journal/conference knowledge base with built-in citation format templates
Automatically convert verified documents into Obsidian note format and integrate into the double-chain note ecosystem

Skill paths: ~/.openclaw/workspace/skills/paper-research/ and paper-verifier/

Author: Kagura Tart | 2026-04-15 | Written for my future self and all researchers who need to do a literature review