Why This Open‑Source Claude Code Pipeline Has Earned 6.4k Stars for AI‑Powered Paper Writing

ARS Overview

academic-research-skills (ARS) is a Claude Code skill set that implements a full academic research pipeline through four sequential skills: Deep Research, Academic Paper, Academic Paper Reviewer, and Academic Pipeline.

1. Deep Research (13 agents)

Agents perform literature search, problem formulation, methodology design, and systematic PRISMA reviews. One agent queries the Semantic Scholar API and validates each reference using Levenshtein similarity > 0.70. Additional agents include a Socratic mentor for guided reasoning and a Devil’s Advocate that challenges early ideas.

2. Academic Paper (12 agents)

The writing team generates outlines, constructs arguments, drafts text, produces bilingual abstracts, visualizes figures, and converts citations to APA 7.0 or IEEE PDF. A style‑calibration function learns the user’s past writing style.

3. Academic Paper Reviewer (7 agents)

The reviewer simulates a journal review: an editor‑in‑chief leads three domain reviewers plus a Devil’s Advocate. Scores are 0‑100; ≥80 = accept, 65‑79 = minor revision, 50‑64 = major revision, <50 = reject. The team also outputs a detailed revision roadmap.

4. Academic Pipeline

The pipeline orchestrates the three teams into a ten‑stage linear process, with optional entry points (e.g., Stage 2.5 for an existing draft or Stage 4 after reviewer comments).

Key Design Innovations

Citation Verification : Each cited paper is cross‑checked via the Semantic Scholar API; fuzzy matching uses Levenshtein similarity > 0.70 to reject hallucinated citations.

Integrity Gates : Stages 2.5 and 4.5 run a 7‑item AI‑failure checklist derived from a 2026 Nature study (covers citation hallucination, data fabrication, methodological falsification, etc.). In a real‑paper test the system flagged 15 fabricated citations and 3 statistical errors.

Anti‑Flattery Protocol : A Devil’s Advocate scores rebuttals 1‑5; if the score is below 4 the writing AI cannot concede, preventing overly conciliatory responses.

Three‑Layer Data Isolation : Layer 1 – raw, untrusted input; Layer 2 – verified output; Layer 3 – gold‑standard scoring data and reference answers, never exposed to the writing agents. This mirrors the isolation model from Anthropic’s w2s‑researcher work.

Reproducibility Lock : Each artifact includes a repro_lock file documenting the exact runtime configuration, with a disclaimer that LLM outputs are not byte‑level reproducible.

Cost and Resource Consumption

A 1.5 k‑word paper costs roughly $4‑6 to process. The full ten‑stage pipeline consumes >200 k input tokens and 100 k output tokens. Recommended model is Claude Opus 4.7 with the Max subscription ($100 or $200 per month).

Installation

/plugin marketplace add Imbad0202/academic-research-skills

/plugin install academic-research-skills

Verify the installation with:

/ars-plan

Alternatively, upload SKILL.md to the Claude.ai knowledge base for a single‑agent experience.

Repository

Source code and documentation: https://github.com/Imbad0202/academic-research-skills