6.4k Stars! An Open-Source Pipeline for Writing Papers with Claude Code

An open-source project called academic-research-skills (ARS) has garnered 6.4k stars on GitHub, offering a complete pipeline for writing academic papers using Claude Code. Developed by Edward Cheng-I Wu (Wu Zhengyi) from Taiwan, China, the project aims to systematically enhance the reliability and efficiency of AI-assisted academic research through multi-agent collaboration.

ARS's core architecture consists of four skills: Deep Research, Academic Paper, Academic Paper Reviewer, and Academic Pipeline. Deep Research assembles a 13-agent research team responsible for literature review, research question formulation, methodology design, and systematic PRISMA reviews. It includes a literature tracing agent that calls the Semantic Scholar API to verify each citation's existence, a Socratic tutor agent for guided ideation, and a devil's advocate agent to challenge assumptions.

Academic Paper is a 12-agent writing team covering outline design, argument construction, draft writing, bilingual abstract generation, chart visualization, and citation format conversion. A standout feature is style calibration, where the AI learns the user's previous writing style to produce output that feels personal rather than generic. Output formats include Markdown, DOCX, and LaTeX, which can be compiled into APA 7.0 or IEEE format PDFs.

Academic Paper Reviewer is a 7-agent review team simulating real journal peer review. Led by an Editor-in-Chief (EIC) with three domain reviewers and a devil's advocate, it scores papers from multiple dimensions using a 0-100 scale: ≥80 accept, 65-79 minor revision, 50-64 major revision, <50 reject. The team also produces a detailed revision roadmap.

Academic Pipeline orchestrates the three teams into a 10-stage workflow: research, writing, integrity check, peer review, revision, final check, publication preparation, and process summary. Users can join at any stage—for instance, starting from Stage 2.5 (integrity check) if they already have a draft, or jumping directly to Stage 4 (revision) when review comments are received.

ARS incorporates several key mechanisms to prevent AI from sabotaging academic research. First, citation verification: during Deep Research, every reference is checked via Semantic Scholar API using Levenshtein similarity (threshold ≥0.70) to avoid hallucinated citations. Second, integrity gates: at Stages 2.5 and 4.5, mandatory checks run a 7-item AI failure mode checklist derived from a 2026 Nature study on autonomous AI research, covering citation hallucination, data fabrication, and methodological fraud. Any issue flagged as SUSPECTED at Stage 2.5 must be resolved to CLEAR at Stage 4.5 or manually overridden with a record. In real tests, this caught 15 fake citations and 3 statistical errors in one paper.

Third, an anti-sycophancy protocol: the devil's advocate's objections are scored 1-5; if below 4, the writing team is not allowed to concede, preventing the AI from pleasing the user. Additionally, the intensity of criticism must be maintained across revisions—if the first review heavily criticizes the methodology, the reviewer cannot become gentler after revision. Score trajectories are tracked, and any dimension's score drop is flagged as a regression, analogous to not introducing new bugs in software engineering.

Fourth, three-layer data isolation: Layer 1 contains raw input (untrustworthy), Layer 2 holds integrity-verified content, and Layer 3 includes scoring criteria, reference answers, and gold-standard data (never exposed to the writing AI). The writing and review teams are invoked independently with stage boundaries. The writing AI receives only natural language feedback from the review AI (e.g., "Chapter 2 has a logic gap, suggest adding comparative experiments") but not the scoring rubric. This design is inspired by Anthropic's w2s-researcher study, which found that when AI reads label data, it may optimize surface features rather than genuinely generalize.

Finally, honest documentation: each artifact is accompanied by a repro_lock file recording the full runtime configuration. However, it includes a mandatory disclaimer that LLM outputs are not byte-level reproducible, model weights may update without changing model IDs, and external APIs return different data daily. The file is a configuration document, not a replay guarantee.

Installation is straightforward: users of Claude Code can install ARS with two commands. Alternatively, uploading the SKILL.md file to a claude.ai project knowledge base enables a single-agent version in the browser (limited functionality). The project supports Traditional Chinese and English.

Regarding cost, the developer recommends Claude Opus 4.7 with a Max subscription plan. A full 10-stage run consumes over 200k input tokens and 100k output tokens. Max plans cost $100 or $200 per month. For a 15,000-character paper, the estimated cost per complete run is approximately $4-6.

Project link: https://github.com/Imbad0202/academic-research-skills