If you use Claude Code for academic research, there is an open-source project worth two minutes of your time to install. academic-research-skills (ARS) is a Claude Code skill pack with 6.4k GitHub stars, containing 4 skills that cover the full paper workflow: research, writing, review, and finalization. Most importantly, it does not simply have AI write your paper — it uses a systematic set of mechanisms to keep the AI from making mistakes, flattering you, or fabricating.

What Each of the 4 Skills Does

ARS's core architecture consists of 4 skills that fit together into a complete pipeline from topic selection to submission.

Deep Research (a research team of 13 agents)

Handles literature review, research question construction, and methodology design, and can also write PRISMA reviews. The team includes a literature-tracing agent (calls the Semantic Scholar API to verify that citations actually exist), a Socratic mentor agent (guides you through dialogue to clarify your thinking), and a devil's advocate agent (specifically pokes holes to prevent tunnel vision).

Academic Paper (a writing team of 12 agents)

Covers everything from outline design, argument construction, and draft writing to bilingual abstract generation, figure and chart visualization, and citation format conversion. Its signature feature is style calibration — the AI learns the writing style of your past work so the output reads more like something you wrote yourself. Output formats include Markdown, DOCX, and LaTeX, and the final result can be compiled into an APA 7.0 or IEEE-format PDF.

Academic Paper Reviewer (a review team of 7 agents)

Simulates a real academic journal review process: an editor-in-chief (EIC) leads three field reviewers plus a devil's advocate, scoring the paper on methodology, disciplinary perspective, interdisciplinary value, and more. Scoring uses a 0-100 quantitative scale: 80 or above accepts, 65-79 minor revisions, 50-64 major revisions, below 50 rejects. It also outputs a detailed revision roadmap.

Academic Pipeline (the workflow orchestrator)

Chains the three teams above into a 10-stage pipeline. From research, writing, integrity checks, peer review, revision, and final checks through to publication prep, every stage has clear deliverables and checkpoints, and you can jump in at any stage.

4 Notable Design Details

1. Citation Verification

The biggest hazard when AI writes papers is hallucinated citations. ARS builds a citation verification mechanism into the Deep Research stage: every reference must pass an existence check against the Semantic Scholar API. It does not simply match titles — it uses a Levenshtein similarity algorithm for fuzzy matching, and only passes above a 0.70 threshold. In real-world testing, this mechanism caught 15 fabricated citations and 3 statistical errors in a single genuine paper.

2. Integrity Gates

At Stage 2.5 and Stage 4.5 of the pipeline there are two integrity gates that cannot be skipped, and each runs a 7-item AI failure mode checklist. The checklist comes from a 2026 Nature study on fully autonomous AI research, covering citation hallucination, data fabrication, methodology fraud, and similar failure modes.

Any issue flagged as SUSPECTED at 2.5 must turn CLEAR at 4.5, or be manually overridden by a human with the override recorded. The design logic: turn "I trust the AI not to make mistakes" into "I require the AI to prove it made no mistakes."

3. Anti-Sycophancy Protocol

Most AI tools have a tendency to flatter the user. ARS builds an anti-sycophancy mechanism specifically into the review stage: the devil's advocate's rebuttals are scored 1-5, and if one scores below 4, the writing team is not allowed to concede. The AI cannot cave just to seem agreeable. At the same time, attack intensity must hold steady through revision — if the first round of review tears the methodology apart, the reviewers cannot suddenly turn gentle after the author revises.

4. Three-Layer Data Isolation

• Layer 1: raw input, untrusted by default
• Layer 2: outputs that have passed integrity verification
• Layer 3: scoring rubrics and gold-standard data — this layer must never appear in the writing AI's context

The writing AI only receives the review AI's natural-language feedback (for example, "the argument in Chapter 2 jumps — add a comparison experiment"), but it never sees the original scoring rubric.

Installation

Option 1: Claude Code Plugin (Recommended)

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

Verify the installation:

/ars-plan

Then describe your paper topic, and ARS will start a Socratic dialogue to help you structure the paper.

Test a single command:

/ars-lit-review "your research topic"

Option 2: claude.ai Web

Simply upload SKILL.md to a claude.ai project's knowledge base. No Claude Code installation needed — open a browser and it works. Note that this mode does not support multi-agent parallelism; it is a single-agent version suited to light experimentation.

Cost Reference

• Recommended setup: Claude Opus 4.7 paired with a Max subscription plan
• A full 10-stage run can consume more than 200K input tokens and 100K output tokens in a single pass
• A 15,000-word paper costs about $4-6 for the full run
• Max subscription plans come in two tiers: $100 or $200 per month
• Using an individual sub-module (say, only the literature review) costs far less

Who It's For

• Graduate students and researchers who need to use AI systematically for paper writing
• Scholars concerned about citation hallucination and sycophancy in AI-assisted writing
• Teams that want an auditable, traceable AI research workflow

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