Academic Research Skills for Claude Code

Academic Research Skills (ARS) is an open-source suite of Claude Code skills built and maintained by Cheng-I Wu. It covers the full academic research pipeline — from deep literature review through paper writing, peer review simulation, and final publication — using multi-agent orchestration inside Claude Code. The project is licensed under CC BY-NC 4.0 and is freely available on GitHub.

What It Is

ARS is a plugin for Claude Code (Anthropic's agentic coding and research CLI) that adds four specialized skill modules: Deep Research, Academic Paper, Academic Paper Reviewer, and Academic Pipeline. Each skill deploys a team of AI agents — ranging from 7 to 13 per skill — that handle distinct tasks such as reference verification, argument construction, peer review simulation, and integrity checking. The tool is explicitly designed as a human-in-the-loop system: the README states "AI is your copilot, not the pilot," and every pipeline stage requires explicit user confirmation before proceeding.

Architecture and Pipeline

The four skills form a 10-stage orchestrated pipeline:

• Deep Research (v2.8) — 13-agent team with 7 modes including full, quick, PRISMA systematic review, Socratic guided, fact-check, lit-review, and review. Includes Semantic Scholar API verification and cross-model divergence analysis.
• Academic Paper (v3.0) — 12-agent writing pipeline with 10 modes covering full drafting, outline-only, revision coaching, citation conversion, LaTeX hardening, VLM figure verification, and AI disclosure statement generation.
• Academic Paper Reviewer (v1.8) — 7-agent multi-perspective review with 0–100 quality rubrics, EIC + 3 dynamic reviewers + Devil's Advocate, and an optional calibration mode that measures its own false-negative/false-positive rates against a user-supplied gold set.
• Academic Pipeline (v3.7) — 10-stage orchestrator with mandatory integrity gates at Stages 2.5 and 4.5, R&R Traceability Matrix, Collaboration Depth Observer, and a Stage 6 Process Summary with 6-dimension Collaboration Quality Evaluation.

Integrity and Anti-Hallucination Design

A central design concern is citation and data hallucination. The pipeline includes two mandatory integrity verification stages (2.5 and 4.5) that cannot be skipped. The README documents a post-publication audit of a real pipeline run that found 21 out of 68 references had issues — a 31% error rate — after three rounds of integrity checks, which motivated the v2.7 anti-hallucination overhaul. The integrity agent uses a three-outcome taxonomy (VERIFIED / NOT_FOUND / MISMATCH) with mandatory WebSearch audit trails and Semantic Scholar API cross-checks.

Anti-Sycophancy and Dialogue Health

Version 3.0 introduced structural countermeasures against known AI behavioral failure modes. The Devil's Advocate agent must score every rebuttal on a 1–5 scale before conceding; concession is only permitted at score ≥4. The Socratic Mentor classifies user intent as exploratory vs. goal-oriented every three turns, disabling auto-convergence in exploratory mode. A silent Dialogue Health Indicator self-assesses every five turns for persistent agreement, conflict avoidance, and premature convergence, auto-injecting challenges when patterns are detected.

Update: v3.7.0 — Claude Code Plugin Packaging (2026-05-05)

The most recent release, v3.7.0, upgrades ARS from a manual git clone + symlink installation to a one-line Claude Code plugin install:

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

This works across Claude Code CLI, VS Code, and JetBrains. The release adds a plugin manifest, 10 slash commands (/ars-plan, /ars-lit-review, etc.) with model routing pinned per command (Opus for architectural/review-depth modes, Sonnet for the rest), 3 plugin-shipped agents as relative symlinks, and a SessionStart announce hook that injects available commands into the LLM's first turn. The traditional symlink flow remains fully supported. No changes were made to the four skill directories, 25 modes, or agent prompts.

Setup and Supported Outputs

Prerequisites are Claude Code (latest), an ANTHROPIC_API_KEY, and optionally Pandoc (for DOCX output) and tectonic + Source Han Serif TC (for APA 7.0 PDF compilation). Supported citation formats include APA 7.0, Chicago, MLA, IEEE, and Vancouver. Supported paper structures include IMRaD, Thematic Literature Review, Theoretical Analysis, Case Study, Policy Brief, and Conference Paper. The tool defaults to Traditional Chinese when the user writes in Chinese and English otherwise.