Academic Research Skills for Claude Code (ARS)

  • Author/Source: Edward Cheng-I Wu (吳政宜), GitHub

  • Original: https://github.com/Imbad0202/academic-research-skills

  • Key Ideas

  • ARS is a Claude Code plugin suite of ~25 modes across four skills covering the full pipeline from research question to peer-reviewed paper: Deep Research (13 agents, 7 modes), Academic Paper (12 agents, 10 modes), Academic Paper Reviewer (7 agents, 6 modes), and Academic Pipeline (10-stage orchestrator).
  • Built explicitly as human-in-the-loop, motivated by Lu et al. (2026, Nature) on autonomous AI Scientist failure modes — implementation bugs, hallucinated results, shortcut reliance, frame-lock, methodology fabrication, citation hallucination.
  • Integrity gates at Stages 2.5 and 4.5 run a 7-mode blocking checklist for AI failure modes; cannot be skipped. Tier 0 Semantic Scholar API verification catches fabricated references before review.
  • Anti-sycophancy mechanisms: Devil's Advocate Concession Threshold Protocol (must score rebuttals 1–5 before conceding; concession only at ≥4); Socratic Mentor intent detection (exploratory vs. goal-oriented modes); dialogue health indicator self-checks every 5 turns.
  • Style Calibration learns voice from 3+ past papers; Writing Quality Check flags AI-tell patterns (25 high-frequency terms, em-dash overuse, throat-clearing openers, Rule of Three, uniform paragraph rhythm). Stated framing: better writing, not detection evasion.
  • Cost estimate ~$4–6 for a 15k-word paper through the full pipeline. Output: Markdown → DOCX (Pandoc) → LaTeX (APA 7.0 apa7 class / IEEE / Chicago) → PDF (tectonic).

  • One-line plugin install: /plugin marketplace add Imbad0202/academic-research-skills then /plugin install academic-research-skills. Supports English and Traditional Chinese natively; intent-based mode activation works in any language.

  • Summary

ARS is one of the most architecturally elaborate Claude Code skill bundles for academic writing currently public, drawing on three specific papers in its design: Lu et al. (2026, Nature 651:914–919) on the AI Scientist's failure modes, PaperOrchestra (Song et al., 2026, Google) for Semantic Scholar verification and anti-leakage, and Wang & Zhang (2026, IJETHE 23:11) on collaboration depth. The pipeline runs ten stages (Research → Plan → Write → Integrity 2.5 → Review → Re-review → Revise → Integrity 4.5 → Finalize → Process Summary) with mandatory user-confirmation checkpoints and an append-only "Material Passport" carrying provenance across stages.

The author treats peer review as a multi-agent panel: an Editor-in-Chief, three dynamic reviewers, and a Devil's Advocate score the manuscript on 0–100 rubrics mapped to standard decisions (≥80 Accept, 65–79 Minor, 50–64 Major, <50 Reject). A Calibration mode measures the reviewer's own FNR/FPR against a user-supplied gold set. The release notes are remarkable for their forensic honesty: v2.7 records a post-publication WebSearch audit of 68 references that found a 31% error rate that had survived three rounds of integrity checks, which became the case for adding external verification by default. v3.7 adds plugin packaging, model routing (opus for full pipeline and revision-coach modes, sonnet elsewhere, no Haiku), and slash-command shortcuts (/ars-plan, /ars-lit-review, etc.).

  • Relevance to Economics Research

ARS is one of the clearest worked examples of a research workflow that takes AI failure modes (especially citation hallucination, frame-lock, and sycophancy) seriously enough to engineer specific countermeasures into the toolchain. For economists, the directly useful pieces are: (1) the integrity-gate pattern (verify references with an API, not the LLM's memory) which generalizes to any empirical replication workflow; (2) the Devil's Advocate concession protocol, which is a concrete answer to the recurring complaint that LLM reviewers cave under pushback; (3) the journal taxonomy in top_journals_by_field.md (incorporating the AIS Senior Scholars' Basket of 11); and (4) the Style Calibration / Writing Quality Check pair as a template for keeping AI-assisted prose from drifting toward generic LLM cadence. The pipeline is heavily oriented toward IS/management venues but adaptable to economics with custom prompts.