跨学科端到端学术研究 AI 技能包 — 从构思评估到终稿提交，覆盖全学科
Cross-disciplinary end-to-end academic research skill pack for Claude Code — from idea evaluation to camera-ready submission. Supports 27 sub-commands, 28+ scientific databases, 12+ venue templates, and cross-platform LaTeX compilation (macOS, Linux, Windows).
跨学科学术研究技能包 — 从文献综述到 camera-ready 提交。支持 27 个子命令、28+ 科学数据库、12+ 顶级会议/期刊模板，跨平台 LaTeX 编译。
English · 中文 · MIT License
Academic research has a "last-mile problem": you've read the papers, you have the idea, but turning it into a submission-ready manuscript still requires navigating dozens of workflows — literature search, citation management, LaTeX formatting, integrity verification, De-AIGC, and peer review simulation.
Each of these steps has pitfalls. AI can generate plausible-sounding but fabricated citations. LLMs are unreliable with numbers. And without verification, errors compound silently through drafts.
This skill pack solves this by encoding the entire research pipeline into 20 specialized agents, enforced by 7 non-negotiable Iron Rules derived from Lu et al. (2026, Nature) and Zhao et al. (2026, arXiv:2605.07723).
Our vision: AI suggests, human decides. AI executes, human verifies. The researcher always remains the PI.
┌───────────────────────────────────────────────────────────┐
│ 12 Slash Commands (/ar-*) │
│ + 27 Sub-Commands (/hbe-academic) │
├───────────────────────────────────────────────────────────┤
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Research │ │ Writing │ │ Review │ │ Pipeline │ │
│ │ 5 agents │ │ 5 agents │ │ 4 agents │ │ 6 agents │ │
│ └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ │
│ └─────────────┴────────────┴─────────────┘ │
│ │ │
│ ┌──────────────────────┴──────────────────────────────┐ │
│ │ Material Passport (ar-passport.json) │ │
│ │ Cross-session state · Stage tracking · Audit trail │ │
│ └──────────────────────┬──────────────────────────────┘ │
│ │ │
│ ┌──────────────────────┴──────────────────────────────┐ │
│ │ 7 Iron Rules · 2 Integrity Gates · De-AIGC │ │
│ │ Non-negotiable academic integrity enforcement │ │
│ └──────────────────────┬──────────────────────────────┘ │
│ │ │
│ ┌──────────────────────┴──────────────────────────────┐ │
│ │ 28+ Scientific Databases (MCP) · 140+ Tool Guides │ │
│ │ arXiv · PubMed · OpenAlex · Semantic Scholar · FRED │ │
│ └──────────────────────┬──────────────────────────────┘ │
│ │ │
│ ┌──────────────────────┴──────────────────────────────┐ │
│ │ 12+ LaTeX Templates · Cross-Platform Compilation │ │
│ │ NeurIPS · ICML · ICLR · ACL · IEEE · ACM · ... │ │
│ │ macOS (MacTeX) · Linux (TeXLive) · Windows (MiKTeX)│ │
│ └─────────────────────────────────────────────────────┘ │
└───────────────────────────────────────────────────────────┘
Stage 0: Idea Eval → Stage 1: Plan → Stage 2: Draft
→ Stage 2.5: INTEGRITY GATE (mandatory) ←
Stage 3: Review → Stage 3.5: Revision → Stage 4: Finalize
→ Stage 4.5: INTEGRITY GATE (mandatory) ←
Stage 5: Submit
The two integrity gates are non-negotiable. The pipeline cannot advance without passing them. This prevents unverified claims, fabricated citations, and AI-typical writing from reaching the final manuscript.
| # | Rule | Category |
|---|---|---|
| 1 | Never fabricate a citation — every reference must be verified against a real database | Integrity |
| 2 | Never trust unverified AI numbers — re-calculate, cross-reference, or reproduce | Accuracy |
| 3 | Every claim needs evidence — citation, experimental data, or labeled conjecture | Rigor |
| 4 | Integrity gates cannot be skipped — stages 2.5 and 4.5 are mandatory | Process |
| 5 | The researcher is the PI — AI suggests, human decides | Authority |
| 6 | Five-step citation verification — existence, accuracy, metadata, BibTeX, consistency | Verification |
| 7 | De-AIGC checks before submission — five dimensions, all ≥ 7 | Quality |
The ar-passport.json manages cross-session state for long-running research projects:
- Pipeline stage (0–5) with full history — resume exactly where you left off
- Integrity gate verdicts — auditable, immutable after passage
- Artifacts — every file produced, with hashes and timestamps
- Style profile — 5-dimension calibrated writing voice
- Bibliography state — verified count, pending references
- Collaboration log — every human-AI interaction recorded
Learns and preserves the researcher's unique writing voice:
| Dimension | Measures | Example Values |
|---|---|---|
| Sentence Length | short / mixed / long | CS: short · Humanities: long |
| Hedge Frequency | low / moderate / high | Physics: low · Social Science: high |
| Passive Ratio | 0.0–1.0 float | Economics: 0.10 · Medicine: 0.40 |
| Jargon Density | low / medium / high | Math: high · Education: low |
| Paragraph Structure | topic-evidence / claim-data / assertion-elaboration | Engineering: claim-data-warrant |
Every claim in the manuscript is independently verified:
| Level | Method | When Used |
|---|---|---|
| Primary | Go to the original source (DOI, arXiv, PDF) and verify | Every cited claim |
| Secondary | Independent re-calculation or cross-reference with a different source | Statistical and experimental claims |
| Tertiary | A fresh Claude session (blind to original context) reviews the evidence chain | Integrity gates, high-stakes claims |
| Group | Agents | Role |
|---|---|---|
| Research | literature_search, synthesis, bibliography, socratic_mentor, fact_check |
28+ database systematic search, multi-paper synthesis, citation lifecycle, guided exploration, claim verification |
| Writing | draft_writer, figure_design, de_aigc, citation_compliance, visualization |
IMRAD drafting, 3 load-bearing figures, De-AIGC 5-dimension assessment, venue-specific formatting, scientific plots |
| Review | peer_reviewer, devils_advocate, integrity_verification, editorial_synthesizer |
7-dimension simulated review, active claim challenging, 4-layer integrity pipeline, multi-reviewer feedback synthesis |
| Pipeline | orchestrator, compliance, collaboration_observer, style_calibration, claim_audit, deep_reader |
8-stage management, venue format compliance, human-AI depth monitoring, voice calibration, claim-faithfulness audit, four-pass paper reading |
- Agents do NOT call other agents directly — only slash commands (
/ar-*) orchestrate composition - Every agent reads the Material Passport for context and writes back its contribution
- Iron Rules apply to all agents without exception
- All agents produce bilingual (EN/CN) content when the user writes in Chinese
| Command | Purpose |
|---|---|
/ar-full |
Complete 8-stage pipeline (idea → camera-ready) |
/ar-plan |
Socratic research planning |
/ar-lit-review |
Systematic literature review |
/ar-reviewer |
Multi-perspective peer review simulation |
/ar-abstract |
Abstract writing and polishing |
/ar-outline |
Paper outline generation |
/ar-revision |
Revision and rebuttal workflow |
/ar-revision-coach |
Interactive revision coaching |
/ar-citation-check |
Citation compliance verification |
/ar-format-convert |
Cross-venue format conversion |
/ar-disclosure |
AI usage disclosure generation |
/ar-deep-read |
Four-pass deep paper reading |
Per-package deep dive guides covering the research lifecycle:
Data Acquisition → Processing → EDA → Modeling → Visualization → Writing
│ │ │ │ │ │
gget pandas scanpy PyTorch matplotlib LaTeX
Biopython polars seaborn scikit-learn pgfplots natbib
pysam dask plotly transformers TikZ biblatex
pydicom vaex seaborn econml Inkscape docx
...140+ ... ... ... ... ...
See references/tool-registry.md for the full catalog, and references/tools/ for per-package deep guides.
/plugin marketplace add ViewWay/academic-research-skill
/plugin install academic-research-skill
git clone https://github.com/ViewWay/academic-research-skill.git ~/github/academic-research-skill
ln -s ~/github/academic-research-skill ~/.claude/skills/academic-researchWorks standalone or as an HBE (Hermes-by-Everything's) add-on skill.
# Check LaTeX environment
/hbe-academic check-env
# Start a new paper project
/hbe-academic paper --venue neurips --title "Your Paper Title"
# Literature review
/hbe-academic lit-review --topic "your research topic"
# Deep read a paper
/ar-deep-read
# Run full pipeline
/ar-full
# Compile LaTeX
/hbe-academic compile --engine xelatex --template neurips| Venue | Type | Template | Compiler |
|---|---|---|---|
| NeurIPS | Conference | ✅ | pdflatex/xelatex |
| ICML | Conference | ✅ | pdflatex/xelatex |
| ICLR | Conference | ✅ | pdflatex/xelatex |
| ACL/EMNLP | Conference | ✅ | pdflatex/xelatex |
| AAAI | Conference | ✅ | pdflatex/xelatex |
| IEEE (Trans/Conf) | Journal/Conf | ✅ | pdflatex |
| Springer (LNCS) | Conference | ✅ | pdflatex |
| Elsevier | Journal | ✅ | pdflatex |
| ACM | Conference | ✅ | pdflatex |
| APS (Phys Rev) | Journal | ✅ | pdflatex |
| Chinese Thesis | Thesis | ✅ | xelatex |
| Beamer Slides | Presentation | ✅ | pdflatex/xelatex |
academic-research-skill/
├── SKILL.md # Main skill definition (686 lines)
├── _meta.json # Publishing metadata
├── README.md # This file
├── LICENSE # MIT license
├── .claude-plugin/ # Plugin manifest
│ ├── plugin.json # 20 agents, 12 commands
│ └── marketplace.json # Marketplace listing
├── hooks/ # Session lifecycle hooks
│ ├── hooks.json
│ └── session-start.sh
├── agents/ # 20 specialized agents
│ ├── README.md # Agent composition rules
│ ├── research/ # literature_search, synthesis, bibliography, socratic_mentor, fact_check
│ ├── writing/ # draft_writer, figure_design, de_aigc, citation_compliance, visualization
│ ├── review/ # peer_reviewer, devils_advocate, integrity_verification, editorial_synthesizer
│ └── pipeline/ # orchestrator, compliance, collaboration_observer, style_calibration, claim_audit, deep_reader
├── commands/ # 12 slash command definitions
├── shared/ # Shared protocols and state
│ ├── iron-rules.md # 7 non-negotiable integrity rules
│ ├── anti-patterns.md # Writing/research/pipeline/agent anti-patterns
│ ├── material-passport-schema.md # Cross-session state JSON schema
│ ├── style-calibration-protocol.md # 5-dimension writing voice calibration
│ ├── collaboration-depth-rubric.md # 4-level human-AI collaboration standard
│ └── cross-model-verification.md # 3-level claim verification protocol
├── tests/ # Test suite
├── references/ # Reference documents (18 guides + 140+ tool guides)
│ ├── writing-guide.md
│ ├── citation-workflow.md
│ ├── latex-environment.md
│ ├── journal-templates-guide.md
│ ├── systematic-review-methodology.md
│ ├── tool-registry.md # 60+ curated tool catalog
│ ├── tools/ # Per-package deep guides (140+ packages)
│ └── ... # See references/ for full list
├── templates/ # LaTeX templates (12+ venues)
├── scripts/ # Helper scripts
│ ├── compile.sh
│ ├── check-environment.sh
│ └── search_arxiv.py
└── workflows/ # Sub-workflow definitions
├── literature-review.md
├── paper-writing.md
├── experiment-design.md
├── rebuttal.md
└── ml-benchmarking.md
This skill pack integrates with:
- Supervisor-Skills — Idea evaluation, Intro six-paragraph model, figure design, pre-submission review
- Awesome-Agent-Skills — StatsPAI causal inference, De-AIGC skills, empirical analysis pipeline
- Overleaf — cloud LaTeX editing via API
- Zotero — reference management
- 28+ Scientific Databases — MCP integration with arXiv, PubMed, OpenAlex, Semantic Scholar, FRED, ADS, SSRN, etc.
Contributions welcome — PRs, issues, and feature requests are all appreciated.
学术研究有一个**"最后一公里问题"**：你读完了文献、有了想法，但把它变成一篇可投稿的稿件，仍需穿越十几个工作流——文献检索、引用管理、LaTeX 排版、诚信验证、降 AIGC、同行评审模拟。
每个步骤都有陷阱。AI 可以生成听起来合理但完全虚构的引用。大语言模型在数字方面不可靠。如果不验证，错误会在草稿间静默累积。
本技能包将整个研究流程编码为 20 个专业智能体，由 7 条不可协商的铁律强制执行。 铁律来源于 Lu 等人（2026, Nature） 和 Zhao 等人（2026, arXiv:2605.07723） 的实证研究。
我们的愿景：AI 建议，人类决定。AI 执行，人类验证。研究者始终是项目主持人。
| 指标 | 数量 |
|---|---|
| 专业智能体 | 20 个（研究 5 + 写作 5 + 审稿 4 + 流程 6） |
| 斜杠命令 | 12 个（/ar-full 到 /ar-deep-read） |
| 子命令 | 27 个（/hbe-academic 扩展能力） |
| 铁律 | 7 条（不可协商的学术诚信底线） |
| 诚信关卡 | 2 个（Stage 2.5 + Stage 4.5，强制执行） |
| 工具指南 | 140+ 个（覆盖 CS/ML、生物、化学、物理、经济、社科） |
| 科学数据库 | 28+ 个（MCP 集成：arXiv、PubMed、OpenAlex 等） |
| 顶级会议模板 | 12+ 个（NeurIPS、ICML、ICLR、ACL、IEEE、ACM 等） |
| 跨平台编译 | macOS / Linux / Windows（MacTeX / TeXLive / MiKTeX） |
| 引用验证步骤 | 5 步（存在性→准确性→元数据→BibTeX→一致性） |
# 插件市场安装（推荐）
/plugin marketplace add ViewWay/academic-research-skill
/plugin install academic-research-skill
# 或独立安装
git clone https://github.com/ViewWay/academic-research-skill.git ~/github/academic-research-skill
ln -s ~/github/academic-research-skill ~/.claude/skills/academic-research/ar-full # 完整 8 阶段流水线
/ar-plan # 苏格拉底式研究规划
/ar-lit-review # 系统性文献综述
/ar-deep-read # 四遍论文精读
/ar-reviewer # 七维度模拟审稿
/ar-citation-check # 引用合规验证如果觉得这个项目对你有帮助，请点亮右上角的 Star！