Imbad0202/academic-research-skills

What it is

Academic Research Skills (ARS) packages five Claude Code skills into one academic workflow: Deep Research (multi-agent research with modes including PRISMA systematic review and Socratic guidance), Academic Paper (multi-agent drafting with style calibration, citation conversion and LaTeX/DOCX/PDF output), Academic Paper Reviewer (multi-perspective review with a Devil's Advocate and re-review), Academic Pipeline (a 10-stage orchestrator) and SR-Screener (blinded study screening with PRISMA 2020 counts). It is aimed at researchers who want AI help on the mechanical parts — reference hunting, formatting, citation checking, consistency checks — while retaining authorship and every substantive decision.

The design is deliberately human-in-the-loop: every pipeline stage stops for user confirmation, and integrity gates at Stage 2.5 and 4.5 run a blocking checklist of known AI research failure modes (hallucinated citations, frame-lock, sycophancy). Skills ship as prompts plus supporting scripts; the project documents its positioning, risks and data flows explicitly (ARCHITECTURE, RISK_REGISTER).

Quick start

Prerequisites: a recent Claude Code install, with ANTHROPIC_API_KEY exported. Pandoc is optional for DOCX output and tectonic plus a CJK serif font for APA 7.0 PDF; Markdown output works without them. Python is only needed for the write-scope guard and a few opt-in commands (SETUP).

Install the plugin from the marketplace inside Claude Code (v3.7.0+):

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

Verify with /ars-plan and describe a paper you are working on; ARS opens a Socratic dialogue about chapter structure. For a single-shot check, run /ars-lit-review "your topic". You can also start a full pipeline in plain chat with prompts like "I want to write a complete research paper" or "Review this paper", then supply the text.

On claude.ai or Claude Desktop paid plans, add the same repository under Customize → Plugins → Add marketplace → Add from a repository. Hooks and plugin agents do not run in plain chat there; per-channel control availability is documented separately (CONTROL_AVAILABILITY). A sibling Codex-native distribution and a Pi wrapper exist for other clients.

Constraints

• License: CC BY-NC 4.0, so non-commercial use only; commercial use needs separate permission (license terms).
• Outputs are not byte-reproducible. The optional repro_lock records configuration, not replay guarantees; LLM outputs vary between runs.
• Verification has a boundary. ARS checks citation existence, claim–source alignment and reported process; it does not establish that experiments were performed, that raw data are authentic, or that results reproduce. Some checks are sampled or LLM-mediated (positioning).
• Reviewer calibration is incomplete. Current live reviews remain NOT_CALIBRATED, and no numerical score maps to Accept/Reject decisions; simulated review is positioned as decision support, not editorial authority.
• Cost: the docs estimate roughly US$3–7 in tokens for a 15k-word paper at 2026-09 list prices before cache discounts (PERFORMANCE).
• Privacy: some data leaves the machine via bibliographic resolvers and optional consent-gated cross-model calls; the documented data-flow page lists what is cached and how to disable each path (DATA_FLOWS).

HotRepo's take

ARS sits in an area that drew immediate skepticism: when it appeared on Hacker News in May 2026, the thread's most-upvoted comment asked for a name for projects shipped "with no obvious validation" (discussion), and a reply argued that evaluating this kind of prompt-heavy tooling now costs more than producing it (comment). That criticism is answerable because the repository is unusually candid about its own limits — the README documents a v2.7 stress test that produced a 31% citation error rate, and the Jan 2026 showcase run used an older pipeline version whose numbers are explicitly not claimed for the current gates (showcase). A separate commenter questioned reliance on tools whose underlying model can silently degrade mid-session (comment). HotRepo tracked consistent daily growth from about 10.9k to 56.3k stars between May and November 2026, with 161–3,184 stars gained per recorded day, and a Python-language daily best rank of 14 on 2026-11-10. The project has since added deterministic citation gates, a claim-audit pass and SR-Screener, so the practical question for adopters is whether those checks are strong enough for their field's standards, not whether the tooling exists at all.