Stop Writing Papers Alone! Claude Code ARS Is the Secret Weapon Top Researchers Use

Stop Writing Papers Alone! Claude Code ARS Is the Secret Weapon Top Researchers Use

What if your next paper took half the time but earned double the citations?

Every academic knows the midnight dread. The cursor blinking on page 47 of a literature review that refuses to cohere. The sinking realization that three "verified" references don't actually support your central claim. The reviewer who demolishes your methodology in Round 2 because you missed a fatal confounder in Round 1. We've been told AI would fix this. Instead, we got chatbots that hallucinate citations, soften under pushback, and write prose so bland it screams "machine-generated."

Here's the dirty secret: the researchers publishing faster aren't working harder—they're augmented.

Enter academic-research-skills—a radical Claude Code plugin that transforms AI from a clumsy ghostwriter into a genuine intellectual sparring partner. With 35+ specialized agents orchestrating everything from Socratic research dialogue to adversarial peer review, ARS doesn't write your paper for you. It handles the grunt work—citation verification, logical consistency checks, statistical error detection—so you can focus on what actually demands your brain: defining the question, choosing the method, interpreting what the data means, and writing the sentence after "I argue that."

This isn't automation. This is augmentation at the level of cognitive architecture. And it's installable in 30 seconds.

What Is Academic Research Skills (ARS)?

Academic Research Skills is a comprehensive, open-source skill suite for Claude Code—Anthropic's agentic coding interface—that implements a complete research → write → review → revise → finalize pipeline. Created by Cheng-I Wu (吳政宜) and actively maintained with contributions from the academic community, ARS represents one of the most sophisticated applications of multi-agent AI systems to scholarly work.

The project emerged from a critical insight: fully autonomous AI research fails catastrophically. Lu et al. (2026, Nature) demonstrated this with The AI Scientist—the first end-to-end autonomous system to publish through blind peer review. Despite this milestone, their Limitations section reads like a horror catalog: implementation bugs, hallucinated results, shortcut reliance, bug-as-insight reframing, methodology fabrication, frame-lock, citation hallucinations. ARS is built on the opposite premise: a human researcher augmented by AI avoids these failure modes better than either alone.

What makes ARS genuinely different from generic AI writing tools? Three structural commitments:

• Human-in-the-loop by design: Every stage requires user confirmation checkpoints; integrity verification gates cannot be skipped
• Anti-sycophancy engineering: The system actively resists the tendency of AI to agree with you too quickly
• Corpus-scale citation integrity: v3.8 implements three-layer citation anchors with optional claim-faithfulness auditing, motivated by Zhao et al.'s (2026) finding of 146,932 hallucinated citations in 2025 alone

The project is licensed under CC BY-NC 4.0 and has evolved through 38+ versions since early 2026, with each release addressing real failure modes discovered through actual paper production.

Key Features That Separate ARS From Everything Else

Deep Research: 13-Agent Investigation Team

ARS deploys a 13-agent research ensemble with seven distinct modes: full investigation, quick briefs, PRISMA systematic reviews, Socratic guided dialogue, fact-checking, literature reviews, and research quality assessment. The Socratic mode features intent detection that distinguishes exploratory thinking from goal-oriented requests—preventing premature convergence when you're still figuring out what questions matter.

Academic Paper: 12-Agent Writing Pipeline

The writing system includes Style Calibration that learns your voice from past papers, Writing Quality Check that catches machine-generated prose patterns, and LaTeX hardening with APA 7.0 apa7 document class compilation. Ten modes cover everything from full paper generation to abstract-only production, citation format conversion (APA/Chicago/MLA/IEEE/Vancouver), and venue-specific AI disclosure statements.

Academic Paper Reviewer: 7-Agent Adversarial Review

This is where ARS gets insane. The reviewer assembles an Editor-in-Chief + 3 dynamic reviewers + Devil's Advocate panel, each scoring on 0–100 rubrics with explicit behavioral indicators. The Devil's Advocate operates under a Concession Threshold Protocol: it must score every rebuttal 1–5 before responding, and can only concede at ≥4. Anti-sycophancy rules prevent the DA from retreating just because you pushed back. Decision mapping: ≥80 Accept, 65–79 Minor Revision, 50–64 Major Revision, <50 Reject.

Academic Pipeline: 10-Stage Orchestrator

The full pipeline coordinates all skills through adaptive checkpoints (FULL/SLIM/MANDATORY), Material Passport provenance tracking, optional repro_lock reproducibility metadata, and cross-model integrity verification. Stage 2.5 and 4.5 integrity gates run a 7-mode blocking checklist that caught 15 fabricated references and 3 statistical errors in a real showcase run.

v3.8 Claim-Faithfulness Audit (L3 Gap Closure)

The latest milestone adds opt-in claim auditing (ARS_CLAIM_AUDIT=1). Every citation carries a three-layer anchor (quote, page, section/paragraph); the audit fetches cited sources and judges whether claims are actually supported. Five HIGH-WARN classes trigger formatter refusal: claim-not-supported, negative-constraint-violation, fabricated-reference, anchorless, and constraint-violation-uncited.

4 Real-World Scenarios Where ARS Transforms Your Workflow

Scenario 1: The Systematic Review That Wouldn't End

You've committed to PRISMA. Six months in, you have 2,400 abstracts, conflicting inclusion criteria, and a supervisor asking why the forest plot still isn't ready. ARS's systematic-review mode deploys risk-of-bias and meta-analysis agents with PRISMA-trAIce compliance checking. The Compliance Agent (v3.4+) runs 17-item PRISMA verification with tier-based blocking—mandatory items halt progress, advisory items surface warnings. The anti-leakage protocol prevents the AI from filling knowledge gaps from parametric memory, flagging [MATERIAL GAP] instead.

Scenario 2: The Revision Round That Broke Your Spirit

Reviewer 2 demolished your framing. Reviewer 3 found a confounder you missed. The EIC wants "major revision." ARS's revision-coach mode parses comments into a structured roadmap, while the re-review mode independently verifies your revision claims against the original critiques via the R&R Traceability Matrix (Schema 11). The matrix adds "Author's Claim" and "Verified?" columns—no more hoping reviewers won't notice you didn't actually address their concern.

Scenario 3: The Citation Hallucination Crisis

Zhao et al.'s audit found 85.3% persistence of hallucinated citations from preprint to publication. ARS's integrity verification agent (v2.7+) eliminated gray-zone classifications—every reference is VERIFIED, NOT_FOUND, or MISMATCH. v3.7.3 added contamination signals (preprint/post-LLM-inflection detection, Semantic Scholar unmatched flags). v3.8's claim audit actually fetches cited passages and judges support. In the showcase post-publication audit, even three rounds of integrity checks missed 21/68 issues that external verification caught—proving why this matters.

Scenario 4: The Cross-Collaboration Coordination Nightmare

Your co-author uses GPT-5.4, you use Claude, the statistician uses R. ARS's cross-model verification protocol (ARS_CROSS_MODEL) runs integrity sample checks and independent Devil's Advocate critique across models. The Collaboration Depth Observer (v3.5+) scores user-AI collaboration on four dimensions—Delegation Intensity, Cognitive Vigilance, Cognitive Reallocation, and Zone Classification—surfacing whether you're actually thinking or just rubber-stamping.

Step-by-Step Installation & Setup Guide

Prerequisites

Before installing ARS, ensure you have:

• Claude Code installed (latest version; plugin packaging requires v3.7.0+)
• ANTHROPIC_API_KEY exported, or set on first claude run
• Optional: Pandoc for DOCX output, tectonic + Source Han Serif TC for APA 7.0 PDF (Markdown^{↗ Smart Converter} output works without either)

Plugin Install (Recommended, 30 Seconds)

The fastest path uses Claude Code's native plugin marketplace:

/plugin marketplace add Imbad0202/academic-research-skills
/plugin install academic-research-skills

Verify the installation by running:

/ars-plan

Describe a paper you're working on—ARS will initiate a Socratic dialogue to map your chapter structure. For a quick single-shot test instead:

/ars-lit-review "your research topic"

Alternative: Traditional Symlink Install

If you prefer manual installation or need customization:

# Clone the repository
git clone https://github.com/Imbad0202/academic-research-skills.git

# Create symlink to Claude Code skills directory
ln -s $(pwd)/academic-research-skills/skills/* ~/.claude/skills/

See docs/SETUP.md for five installation methods including project-local skills, global skills, claude.ai Project integration, and full repo cloning.

Cross-Model Verification Setup (Optional)

For research requiring highest integrity assurance:

# Set environment variable before running Claude Code
export ARS_CROSS_MODEL=1

# Optional: adjust sampling interval for cost control
export ARS_CROSS_MODEL_SAMPLE_INTERVAL=5

This activates independent verification using GPT-5.4 Pro or Gemini 3.1 Pro for integrity sample checks and DA critique. Without this flag, everything works exactly as before—no breaking changes.

Performance Configuration

Per-mode token budgets and full-pipeline cost estimates live in docs/PERFORMANCE.md. Key recommendation: enable Skip Permissions in Claude Code settings for smoother agent handoffs. The full 15,000-word paper pipeline runs approximately $4–6.

REAL Code Examples: How ARS Actually Works

Example 1: Plugin Installation Commands

These are the exact commands from the repository README for getting started:

/plugin marketplace add Imbad0202/academic-research-skills
/plugin install academic-research-skills

After installation, the system loads 10 slash commands mapped to mode triggers. Here's how model routing works—pinned in each command's frontmatter:

# From commands/ars-plan.md frontmatter
model: opus        # Architectural depth for full planning
# vs. commands/ars-lit-review.md
model: sonnet      # Standard research tasks

Why this matters: ARS doesn't use a one-size-fits-all model. The opus model handles architectural planning and revision interpretation where reasoning depth is critical; sonnet covers the other 8 commands for efficiency. The project explicitly excludes Haiku per policy—academic integrity tasks demand sufficient reasoning capacity.

Example 2: Quick Start Conversation Patterns

The README provides exact conversation patterns that trigger different modes through intent detection:

# Start a full research pipeline
You: "I want to write a research paper on AI's impact on higher education QA"

# Start with Socratic guidance
You: "Guide my research on AI in educational evaluation"

# Write a paper with guided planning
You: "Guide me through writing a paper on demographic decline"

# Review an existing paper
You: "Review this paper" (then provide the paper)

# Check pipeline status
You: "status"

The technical insight: These aren't keyword matches. The Socratic and Plan modes use intent-based activation—they detect the meaning of your request, not specific strings. This works in any language without modification. The two-layer architecture separates skill activation (bilingual keywords for confidence) from mode routing (language-agnostic intent signals). When intent is ambiguous, the system defaults to guided modes—safer to explore first than generate prematurely.

Example 3: Deep Research Mode Selection

"Research the impact of AI on higher education"       → full mode
"Give me a quick brief on X"                          → quick mode
"Do a systematic review on X with PRISMA"             → systematic-review mode
"Guide my research on X"                              → socratic mode (guided)
"Fact-check these claims"                             → fact-check mode
"Do a literature review on X"                         → lit-review mode
"Review this paper's research quality"                → review mode

Behind the scenes: Each mode activates a different agent subset from the 13-agent research team. Full mode deploys the complete ensemble; quick mode uses a streamlined 3-agent configuration; systematic-review mode adds PRISMA-specific agents (risk_of_bias, meta_analysis, monitoring). The mode registry is formally specified in MODE_REGISTRY.md with version-anchored entries.

Example 4: Academic Paper Reviewer Decision Mapping

The reviewer implements explicit numerical thresholds with behavioral indicators:

"Review this paper"                                   → full mode (EIC + R1/R2/R3 + Devil's Advocate)
"Quick assessment of this paper"                      → quick mode
"Guide me to improve this paper"                      → guided mode
"Check the methodology"                               → methodology-focus mode
"Verify the revisions"                                → re-review mode
"Calibrate this reviewer against my gold set"         → calibration mode

The 0–100 rubric with decision mapping:

• ≥80: Accept
• 65–79: Minor Revision
• 50–64: Major Revision
• <50: Reject

The calibration mode (v3.2+) measures FNR/FPR against your supplied gold set with 5× ensembling. This addresses a critical gap: most AI "reviewers" have unknown error rates. ARS lets you measure and calibrate reviewer performance against ground truth you trust.

Example 5: Environment Variables for Advanced Control

# Enable claim-faithfulness auditing (v3.8)
export ARS_CLAIM_AUDIT=1

# Enable cross-model verification
export ARS_CROSS_MODEL=1

# Enable passport reset boundaries for long sessions
export ARS_PASSPORT_RESET=1

# Enable Socratic reading-check honesty probe
export ARS_SOCRATIC_READING_PROBE=1

What these do: ARS_CLAIM_AUDIT=1 activates the L3 gap closure—every sampled citation is fetched and judged against its anchor. ARS_PASSPORT_RESET=1 promotes every FULL checkpoint to a context-reset boundary, enabling fresh-session resume from the Material Passport alone. ARS_SOCRATIC_READING_PROBE=1 fires a one-time honesty check when you've cited specific papers, logging decline without penalty. All default to OFF—no behavior change without explicit opt-in.

Advanced Usage & Best Practices

The Anti-Sycophancy Discipline

ARS v3.0 discovered three structural AI limits through actual use: frame-lock (DA attacks arguments but never premises), sycophancy under pushback (conceding too quickly when challenged), and intent misdetection (premature convergence during exploration). The fixes are now protocolized:

• Concession Threshold Protocol: DA scores rebuttals 1–5; concession only at ≥4 with evidence
• No consecutive concessions: Forces sustained engagement with critique
• Exploratory mode: Disables auto-convergence, raises max rounds to 60, prohibits "want me to summarize?" prompts

Pro tip: When the Dialogue Health Indicator (silent every 5 turns) detects persistent agreement, it auto-injects challenging questions. You won't see this happening—it's invisible to prevent gaming—but logs are available for post-session review.

Style Calibration Without Deception

Provide 3+ past papers and ARS learns your voice—sentence rhythm, vocabulary preferences, citation integration style. Critical priority system: discipline norms (hard) > journal conventions (strong) > personal style (soft). This isn't humanization or detection evasion. The Writing Quality Check explicitly catches 25 AI high-frequency terms, em dash overuse (≤3), throat-clearing openers, and structural patterns like uniform paragraph lengths that signal machine generation.

Material Passport as Provenance Ledger

The Material Passport (Schema 9+) is ARS's secret weapon for reproducibility. It tracks:

• literature_corpus[]: Your pre-screened literature with CSL-JSON metadata
• compliance_history[]: Append-only PRISMA-trAIce and RAISE records
• reset_boundary[]: Hash-verified session boundaries for cross-session resume
• repro_lock: Optional configuration snapshot (not byte-replay guarantee—LLM outputs aren't deterministic)

Use resume_from_passport=<hash> to continue in a fresh Claude Code session with full context.

Comparison: ARS vs. Generic AI Writing Tools

The fundamental difference: generic tools optimize for output volume; ARS optimizes for intellectual integrity.

FAQ: What Researchers Actually Ask

Is ARS free to use?

Yes. The project is licensed under CC BY-NC 4.0—free for non-commercial academic use with attribution. You need your own Anthropic API key for Claude Code usage; ARS itself has no subscription fees.

Will this write my paper for me?

Explicitly no. The README states: "AI is your copilot, not the pilot." ARS handles grunt work (reference hunting, citation formatting, data verification, logical consistency) so you focus on defining questions, choosing methods, interpreting data, and crafting arguments. Unlike humanizers, it doesn't help you hide AI use—it helps you write better.

How does ARS prevent citation hallucinations?

Multiple layers: Semantic Scholar API verification with Levenshtein title matching ≥0.70, DOI mismatch detection, mandatory WebSearch audit trails, three-layer citation anchors (quote/page/section), and optional claim-faithfulness auditing that fetches cited passages and judges support. The showcase found 21/68 issues in post-publication audit that earlier rounds missed—proving layered defense is necessary.

Can I use ARS with languages other than English?

Yes. Socratic and Plan modes use intent-based activation that works in any language. Traditional Chinese is fully supported with bilingual abstracts. For other languages, you can add trigger keywords to SKILL.md files to improve matching confidence.

What citation formats are supported?

APA 7.0 (default, including Chinese citation rules), Chicago (Notes & Author-Date), MLA, IEEE, and Vancouver. LaTeX output uses apa7 document class with XeCJK for CJK support.

How much does a full paper cost?

Approximately $4–6 for a 15,000-word paper using recommended Claude Code settings. Per-mode token budgets are documented in docs/PERFORMANCE.md. Cross-model verification increases costs proportionally to sampling interval.

Is my data private?

ARS operates through your local Claude Code installation. No data is sent to external servers beyond your normal Anthropic API calls. The optional Semantic Scholar API calls are rate-throttled (1 req/s, 0.1s with S2_API_KEY).

Conclusion: The Future of Academic Work Is Augmented, Not Automated

The lesson of Lu et al.'s AI Scientist and Zhao et al.'s 111-million-reference audit is clear: fully autonomous AI research inherits catastrophic failure modes at scale. The lesson of ARS is equally clear: structured human-AI collaboration can catch what either misses alone.

This isn't about replacing scholarly judgment. It's about extending it—deploying 35+ specialized agents for the mechanical while you retain the meaningful. The Concession Threshold Protocol, the Dialogue Health Indicator, the three-layer citation anchor, the Material Passport provenance ledger: these aren't features. They're cognitive prosthetics for the modern researcher.

Version 3.8 closes the L3 claim-faithfulness gap. The experiment-agent companion handles empirical validation. The collaboration depth observer ensures you're actually thinking, not just delegating. Every design decision emerges from documented failure modes and real paper production.

Stop writing papers alone. Install ARS in 30 seconds, run /ars-plan, and discover what augmented research actually feels like.

👉 Get academic-research-skills on GitHub

Based on Academic Research Skills by Cheng-I Wu. This work is licensed under CC BY-NC 4.0.