Academic Research Skills Deep Technical Breakdown: How 45+ Agents Collaborate to Complete the Full Workflow from Literature Review to Peer Review

When 45+ Agents Work on One Paper at the Same Time

On May 22, 2026, Academic Research Skills (ARS) reached #3 on GitHub Trending, climbing the chart with +2,579 stars in a single day. It is not just another "AI helps you write papers" tool; it is currently the most complete academic research productivity Agent suite, breaking the entire research pipeline into 4 major skills and 45+ specialized Agents, covering the complete closed loop from research question formation to peer review.

Author Cheng-I Wu's design philosophy: "AI is your co-pilot, not the pilot."

1. System Overview: 4 Major Skills × 45+ Agents

┌─────────────────────────────────────────────────────┐
│                   Academic Research Skills            │
│                      v3.9.4.2                        │
├─────────────────────────────────────────────────────┤
│                                                      │
│  Research Question Formulation (Socratic Dialogue)     │
│         ↓                                            │
│  ┌──────────────────────────────────────────┐       │
│  │  🔬 Skill 1: Deep Research (v2.8)         │       │
│  │  13-Agent research team                   │       │
│  │  Systematic literature review · Meta-analysis · Fact-checking  │       │
│  └──────────────────┬───────────────────────┘       │
│                     ↓                                │
│  ┌──────────────────────────────────────────┐       │
│  │  📝 Skill 2: Academic Paper (v3.0)        │       │
│  │  12-Agent paper-writing pipeline          │       │
│  │  IMRaD structure · citation check · AI disclosure │       │
│  └──────────────────┬───────────────────────┘       │
│                     ↓                                │
│  ┌──────────────────────────────────────────┐       │
│  │  🔍 Skill 3: Paper Reviewer (v1.8)        │       │
│  │  7-Agent peer-review panel                │       │
│  │  Journal-review sim · calibration · editorial calls │       │
│  └──────────────────┬───────────────────────┘       │
│                     ↓                                │
│  ┌──────────────────────────────────────────┐       │
│  │  🔄 Skill 4: Academic Pipeline (v3.7)     │       │
│  │  10-stage master coordinator              │       │
│  │  End-to-end automation · integrity gates · versioning │       │
│  └──────────────────────────────────────────┘       │
│                                                      │
├─────────────────────────────────────────────────────┤
│  Three Major Security Mechanisms                      │
│  ┌───────────────────┐ ┌───────────────────┐ ┌───────────────────┐  │
│  │ Integrity          │ │ Socratic          │ │ Style             │  │
│  │ Gates              │ │ Dialogue          │ │ Calibration       │  │
│  │ Non-skippable      │ │ Guided            │ │ Learns how        │  │
│  │ checkpoints        │ │ questioning       │ │ you write         │  │
│  └───────────────────┘ └───────────────────┘ └───────────────────┘  │
└─────────────────────────────────────────────────────┘

2. Skill 1, Deep Research: 13-Agent Research Team

2.1 Agent Roles and Responsibilities

Deep Research is the core engine of ARS, simulating a complete research team collaboration:

#	Agent Role	Responsibility
1	Research Director	Develops research strategy and assigns tasks
2	Query Formulator	Transforms research questions into searchable queries
3	Literature Searcher	Searches literature across databases (arXiv, PubMed, Semantic Scholar)
4	Relevance Screener	Screens relevant literature (based on title/abstract)
5	Full-Text Reader	Reads full texts in depth and extracts key information
6	Data Extractor	Extracts methods, results, and data in a structured way
7	Quality Assessor	Assesses research quality (risk of bias, methodology)
8	Synthesizer	Synthesizes findings across literature
9	Gap Analyzer	Identifies research gaps
10	Fact Checker	Verifies citations and factual claims
11	Citation Tracker	Tracks citation chains (forward/backward)
12	Bias Detector	Detects selective reporting and publication bias
13	Report Writer	Writes structured systematic review reports

2.2 Workflow

Research Question
    ↓
Query Formulator generates search strategy (PICO framework)
    ↓
Literature Searcher executes multi-database search
    ↓  ↓  ↓
Relevance Screener → Full-Text Reader → Data Extractor
    (process each paper in parallel)
    ↓  ↓  ↓
Quality Assessor + Bias Detector (dual verification)
    ↓
Synthesizer cross-literature synthesis
    ↓
Gap Analyzer identifies gaps
    ↓
Fact Checker + Citation Tracker (citation verification)
    ↓
Report Writer generates systematic review report

2.3 Supported Databases

arXiv
PubMed / PubMed Central
Semantic Scholar
Google Scholar (via web search)
SSRN (Social Sciences)
Custom databases (user-configured)

3. Skill 2, Academic Paper: 12-Agent Writing Pipeline

3.1 Agent Roles

#	Agent	Responsibilities
1	Outline Strategist	Plans the paper structure and argument flow
2	Introduction Writer	Writes the introduction (background + research question)
3	Literature Review Writer	Writes the literature review section
4	Methodology Writer	Writes the methodology section
5	Results Writer	Writes the results section
6	Discussion Writer	Writes the discussion section (interpretation of results + significance)
7	Abstract Writer	Writes the abstract (meets journal word count requirements)
8	Citation Formatter	Formats citations (APA/Chicago/MLA/IEEE/Vancouver)
9	AI Disclosure Writer	Writes the AI usage disclosure statement
10	Flow Checker	Checks the logical flow and fluency of paragraphs
11	Consistency Checker	Ensures consistency of terminology, data, and conclusions
12	Final Polisher	Final language polishing

3.2 Supported Citation Formats

Format	Field	Status
APA 7.0	Psychology, social sciences	✅
Chicago	History, humanities	✅
MLA	Literature, linguistics	✅
IEEE	Engineering, computer science	✅
Vancouver	Medicine, biomedical sciences	✅

3.3 Output Formats

Markdown, native format
DOCX, meets journal submission requirements
PDF, rendered via LaTeX (ACM/IEEE templates)
Bilingual abstract, supports Traditional Chinese and English

4. Skill 3, Paper Reviewer: 7-Agent Review Panel

This may be the most innovative module in ARS, simulating the real peer review process before submission to identify problems in advance.

4.1 The 7 Review Agents

#	Agent	Simulated Role
1	Methodologist	Methodology expert, checks research design and statistical methods
2	Theorist	Theory expert, evaluates theoretical frameworks and contributions
3	Domain Expert	Domain expert, checks the accuracy of domain knowledge
4	Statistician	Statistics expert, verifies data analysis and p-values
5	General Reviewer	General reviewer, checks logic, writing, and overall quality
6	Ethics Reviewer	Ethics reviewer, checks IRB, informed consent, and conflicts of interest
7	Editor-in-Chief	Editor-in-Chief, synthesizes all review comments to make the final decision

4.2 Output Format (Simulated Real Review)

Decision: Minor Revision / Major Revision / Reject

Reviewer 1 (Methodologist):
- Strengths: ...
- Weaknesses: ...
- Specific Comments:
  1. Section 3.2: The sampling method lacks justification...
  2. ...

Reviewer 2 (Theorist):
...

Editor's Decision Letter:
[Simulated editorial decision letter from a real journal]

4.3 Calibration Mode

ARS provides an optional calibration mode: users provide a "gold standard" review set, and the system measures its own false negative rate (FNR) and false positive rate (FPR).

This means you can train the review panel according to your own standards, making it increasingly aligned with the review standards of your target journal.

5. Skill 4, Academic Pipeline: 10-Stage Master Coordinator

5.1 10-Stage Automated Workflow

Stage 1: Research Question Formation ─── Socratic Dialogue
Stage 2: Literature Search Strategy ─── Query Formulation
Stage 2.5: 🔴 Integrity Gate 1 ─ Search Strategy Completeness Verification
Stage 3: Literature Review and Synthesis ─ Deep Research
Stage 4: Paper Outline ──── Outline Generation
Stage 4.5: 🔴 Integrity Gate 2 ─ Outline Logic Verification
Stage 5: Paper Draft ──── Academic Paper
Stage 6: Self-Review ──── Paper Reviewer
Stage 7: Revision ──────── Revision (based on review comments)
Stage 8: Final Check ──── Final Polish + Consistency
Stage 9: Formatting ────── Citation + Template
Stage 10: Delivery ──────── Export (MD/DOCX/PDF)

5.2 Two Mandatory Integrity Gates

The most distinctive design of ARS is its non-skippable Integrity Gates, positioned after key decision points:

Gate	Location	Verification Content
Gate 1	Stage 2.5	Does the search strategy cover the core databases? Are the keywords sufficient?
Gate 2	Stage 4.5	Does the paper structure meet journal requirements? Is the argument chain complete?

The gates include checks for 7 types of AI research failure modes (see ai_research_failure_modes.md), which are heuristics distilled from real-world AI writing failure cases.

6. In-Depth Analysis of the Three Major Safety Mechanisms

6.1 Socratic Dialogue

ARS does not simply execute instructions; it guides researchers to think through questioning:

❌ Generic AI: "I will help you write a paper about X."
✅ ARS: "How is your research question positioned in the existing literature?
        What do you think the main contribution of this study will be?
        What is the rationale behind your methodological choices?"

Intent Detection: ARS distinguishes two interaction modes:

Exploration Mode: The user is thinking and needs guidance.
Goal Mode: The user already has a clear request and needs execution.

6.2 Concession Threshold Protocol

This is an innovative design that addresses a chronic problem in AI writing: under conversational pressure, AI tends to abandon its own position ("You are right").

ARS uses a numerical scoring threshold to prevent AI from collapsing under conversational pressure. The AI accepts revision suggestions only when the criticism exceeds a specific confidence threshold.

6.3 Style Calibration

Before writing begins, ARS asks the user to provide a writing sample:

1. Analyze the user's writing style (sentence patterns, vocabulary, structural preferences)
2. Learn the user's logical organization approach
3. Inject style parameters into all writing Agents

Result: The generated paper sounds like what you wrote, not what AI wrote.

7. Citation and Source Traceability

7.1 Three-Layer Citation Anchors (Locator Infrastructure, v3.7.3+)

Layer	Content	Purpose
L1	Literature source (DOI/URL)	Basic traceability
L2	Page/section locator	Precise citation verification
L3	Claim-citation alignment	Future claim-level audit

7.2 Trust-Chain Frontmatter (v3.7.1+)

Each paper is generated with frontmatter metadata for source provenance, recording the source and reasoning path of every claim. This addresses the pain point of "hallucinated citations."

8. Cost Analysis

8.1 Full Pipeline Cost

Using a standard 15,000-word paper as an example:

Stage	API Calls	Estimated Cost
Deep Research (13 agents)	~80-120	$1.50 - $2.00
Academic Paper (12 agents)	~60-90	$1.20 - $1.80
Paper Reviewer (7 agents)	~40-60	$0.60 - $1.00
Revision + Polish	~30-50	$0.50 - $0.80
Total	~210-320	$3.80 - $5.60

Comparison: Professional academic editing services typically charge $500-$2,000 per paper. ARS costs less than 1% of that.

8.2 Caveats

Using Claude Opus 4.5 (ARS's default model) is more expensive
The system can be configured to use lower-cost models such as DeepSeek v4, but paper quality may decline
Quality validation still requires final oversight by human researchers

9. Comparison with Other Academic AI Tools

Tool	Agent Count	Peer Review	Integrity Gates	Citation Verification	Style Calibration	License
ARS	45+	✅ 7-Agent	✅ Mandatory	✅ Three-tier	✅	CC BY-NC 4.0
ResearchSkills	20+	❌	❌	❌	❌	MIT
Elicit	0 (AI-assisted)	❌	❌	✅	❌	Proprietary
Scite.ai	0 (AI-assisted)	❌	❌	✅	❌	Proprietary
PaperPal	0 (AI-assisted)	❌	❌	❌	❌	Proprietary
Jenni AI	0 (AI-assisted)	❌	❌	❌	❌	Proprietary

ARS is currently the only open-source solution that combines a complete research pipeline + peer review simulation + mandatory integrity verification in a single system.

10. Implications for Junze Think Tank

10.1 Direct Application Scenarios

Our business heavily involves structured research and report writing:

Business Scenario	ARS Module	Room for Improvement
Hong Kong equity due diligence research	Deep Research	Adapt to HKEX data sources (DI + HKEXnews)
DCF valuation report	Academic Paper	Adapt to financial data formats
Investment proposal	Academic Pipeline	Add a financial completeness gate
Industry research report	Deep Research + Paper	Configure industry-specific Agents
Compliance review	Paper Reviewer	Add a regulatory compliance review Agent

10.2 Strategic Recommendations

Learn from its architecture design: The collaboration model of 45+ Agents can be directly applied to our research system
Adapt to business scenarios: Redefine the Agent roles in ARS as "Industry Analyst," "Financial Analyst," "Compliance Reviewer," etc.
Add an HKEX data layer: Integrate our DI, HKEXnews, and CloakBrowser as data sources for ARS
Establish research integrity gates: Draw on the Integrity Gates design to prevent data estimation errors in AK-SDD reports

10.3 Constraints

CC BY-NC 4.0 license: Non-commercial use only; commercial use requires separate negotiation
Claude Code dependency: Requires a Claude subscription (or simulation via Sub2API)
Academic orientation: Currently designed for academic paper scenarios and needs to be adapted for commercial research scenarios

11. Conclusion

Academic Research Skills represents the highest standard in AI-assisted academic writing. Its core contribution is not "stronger writing ability," but systematically solving the trust problem in AI writing through multi-Agent collaboration, mandatory integrity verification, citation traceability, and style calibration.

Advantages	Disadvantages
45+ Agents covering the full workflow	CC BY-NC license restricts commercial use
Mandatory integrity checkpoints ensure quality	Claude model dependency, moderate cost
Three-layer citation traceability	Strong for academic scenarios, needs adaptation for commercial research
Style calibration preserves personal voice	Steeper learning curve (4 independent Skills)
$4-6 per article, exceptionally high cost-effectiveness	Chinese-language support still needs improvement

In one sentence: If ARS were a publishing house, it would simultaneously have a research department, an editorial department, a review department, and a publishing department, and all of this requires only $5 and one Claude Code plugin.

Version: v1.0 · 2026-05-24 · Based on Imbad0202/academic-research-skills v3.9.4.2 (19,200 ⭐)