Claude Code Skills for the Paranoid Academic

What it does

ARS is essentially a sophisticated glue layer for Claude Code: a suite of skills that wires together specialized agent teams to run a 10-stage academic pipeline from literature review through writing, peer review, and final formatting. It handles citation hunting, statistical verification, style calibration, and LaTeX hardening while keeping the human as the decision-maker. The system is explicitly built as a counterargument to fully autonomous AI research tools, positioning the researcher as the pilot and AI as the copilot.

The interesting bit

The project takes citation hallucinations seriously—very seriously. Inspired by corpus-scale audits finding 146,932 hallucinated citations in 2025 alone, ARS implements three-layer citation anchors, an opt-in claim-audit pass that fetches cited sources to verify support, and “integrity gates” at stages 2.5 and 4.5 that run a 7-mode blocking checklist. It even includes a Devil’s Advocate agent in the peer-review stage, because apparently one skeptical reader isn’t enough.

Key highlights

• 10-stage pipeline with adaptive checkpoints and a Material Passport for tracking artifacts across stages
• Deep Research skill with PRISMA systematic review mode and Semantic Scholar API verification
• Writing Quality Check and Style Calibration to catch machine-generated prose patterns and learn the author’s voice
• 7-agent peer review with 0–100 rubrics, cross-model critique, and R&R traceability matrix
• Estimated cost of $4–6 for a full 15,000-word paper pipeline
• Non-commercial CC BY-NC 4.0 license

Caveats

• Requires Claude Code (with Anthropic API key) and optionally Pandoc/tectonic for document conversion; it is not a standalone application
• The reproducibility lockfile is explicitly documented as “configuration documentation, not replay guarantee” since LLM outputs are not byte-reproducible
• Corpus-scale evaluation of ARS’s own anti-hallucination measures remains future work

Verdict

Graduate students and faculty who already use Claude Code and want structured guardrails around literature reviews, citation management, and peer-review simulation will find this saves weeks of grunt work. If you’re looking for a one-click paper generator or balk at per-pipeline API costs, this isn’t it.