Imbad0202/academic-research-skills: a Claude Code plugin for the research to publication pipeline

What academic-research-skills solves, and which researchers it fits

The repository describes itself as a suite of Claude Code skills for academic research, covering the pipeline from research to publication. That scope is narrower than it sounds. The README states directly that the tool will not write your paper for you and frames AI as copilot rather than pilot. The work it claims to absorb is reference hunting, citation formatting, data verification and logical consistency checking, leaving question definition, method choice, data interpretation and the argument sentence to the researcher.

That positioning matters because the alternative design (a system that generates a paper end to end) already exists in the literature the README cites. Lu et al. (2026, Nature 651:914-919) built The AI Scientist, described as the first fully autonomous AI research system to publish a paper through blind peer review at a top-tier ML venue, scoring 6.33/10 against a workshop average of 4.87. The same paper's Limitations section lists implementation bugs, hallucinated results, shortcut reliance, bug-as-insight reframing, methodology fabrication, frame-lock and citation hallucinations. ARS is built on the premise that a human researcher augmented by AI avoids those failure modes better than either alone.

The intended user is therefore someone with a research question already, not someone looking for one. If you are writing a literature review, a thesis chapter or a journal submission and you want the mechanical stages instrumented, the pipeline is aimed at you. If you want the tool to decide what to study, it is not.

How the pipeline works: stages, integrity gates and the citation trust chain

The full pipeline view lives in docs/ARCHITECTURE.md, which the README says supersedes the older sprawling description that used to sit in the README itself. What the README does document is the shape of the flow and the gates inside it.

Stage 2.5 and Stage 4.5 are integrity gates that run a 7-mode blocking checklist, with the checklist itself in academic-pipeline/references/ai_research_failure_modes.md. The reviewer skill offers an opt-in calibration mode that measures its own false negative and false positive rates against a user-supplied gold set. That is a design choice worth noting: rather than asserting the reviewer is accurate, the project ships a way to measure it on your own data.

The citation machinery is layered. Version 3.7.1 added trust-chain frontmatter for source provenance. Version 3.7.3 added locator infrastructure described as three-layer citation anchors, for future claim-level audits, and surfaces advisory risk signals at cite time. Version 3.8 added an opt-in audit pass behind ARS_CLAIM_AUDIT=1 that fetches the cited source against each anchor and judges whether the claim is actually supported. Five HIGH-WARN classes gate-refuse output through the formatter terminal hard gate: claim-not-supported, negative-constraint-violation, fabricated-reference, anchorless and constraint-violation-uncited. Calibration ships as a 20-tuple gold set with FNR below 0.15 and FPR below 0.10 as acceptance thresholds, and the ramp-on plan is deferred to post-calibration evidence per the v3.8 spec.

The motivation is external. Zhao et al. (2026-05) audited 111M references across 2.5M papers on arXiv, bioRxiv, SSRN and PMC, with a conservative estimate of 146,932 hallucinated citations for 2025 alone and an observed mid-2024 inflection. The README is careful here: it says corpus-scale evaluation of ARS itself remains future work. The project cites that audit as design rationale, not as proof that its own gates work at that scale.

Installing academic-research-skills in Claude Code and running your first plan

The README advertises a 30-second install for Claude Code CLI, VS Code and JetBrains on v3.7.0 and later. Two commands from the plugin marketplace do it.

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

After that, the README suggests starting with /ars-plan, which walks through your paper structure via Socratic dialogue. That is the lowest-commitment entry point: it asks questions rather than producing text, which matches the human-in-the-loop framing.

Prerequisites are short. You need Claude Code (the README says latest, because plugin packaging requires recent versions) and ANTHROPIC_API_KEY exported, or set on the first claude run. Pandoc is optional for DOCX output, and tectonic plus Source Han Serif TC is optional for APA 7.0 PDF. Markdown output works without either.

The Python situation is the part people get wrong. The core research, write and review skills need no Python at all, because they are prompt-driven. A real Python interpreter is needed only for the PreToolUse write-scope guard (optional subagent hardening that cleanly no-ops if no real Python is found), plus opt-in features that shell out to Python: revision-patch mode, the submission-package verifier, and the /ars-cache-invalidate, /ars-mark-read and /ars-unmark-read commands. On Windows the README warns that python3 is often a non-functional Microsoft Store placeholder rather than real Python, and points to python.org instead.

Where the gates are advisory, and when this is the wrong tool

The most honest line in the README is about scope of evidence. The project leans on Zhao et al. for motivation and on Ren et al. (2026, Self-Improvements in Modern Agentic Systems: A Survey) for survey-level support, but it states that the survey is cited as design rationale for the human-in-the-loop stance, not as empirical proof that human-in-the-loop pipelines outperform autonomous ones. Corpus-scale evaluation of ARS itself remains future work. Anyone evaluating this on the strength of its citation-integrity claims should read that sentence twice.

The claim-audit pass is opt-in for a reason. ARS_CLAIM_AUDIT=1 fetches cited sources against anchors, which means network access and latency, and the calibration ramp-on is deferred to post-calibration evidence. The 20-tuple gold set with FNR below 0.15 and FPR below 0.10 is a shipped acceptance threshold, not a published result on arbitrary corpora. If your field's citation conventions differ from the ones the gold set encodes, the thresholds tell you nothing about your case.

The tool is also the wrong choice for full automation. If your goal is a system that drafts, reviews and submits without you, the README's own framing rules it out, and the failure-mode checklist in the pipeline references exists precisely because those failures are the ones the project is trying to keep a human in front of. It is equally wrong if you do not use Claude Code: the install path is the plugin marketplace, and the package.json declares a Pi adapter under the pi key with extensions, skills and prompts entries, but the README's install instructions are Claude Code's.

How it differs from general-purpose writing assistants and from autonomous research agents

Two comparisons are worth making, because the project sits between them.

Against a general-purpose writing assistant, the difference is that ARS carries state about citations. The trust-chain frontmatter records source provenance, the three-layer citation anchors give every citation a locator, and the formatter terminal hard gate refuses output on five named HIGH-WARN classes. A generic assistant will happily produce a reference list; it has no terminal gate that can refuse to emit the document. The README is explicit that this is not a humanizer and does not help you hide AI use. Style Calibration learns your voice from past work and Writing Quality Check flags patterns that read as machine-generated. The stated goal is quality, not concealment.

Against The AI Scientist and similar autonomous pipelines, the difference is where the human sits. Lu et al.'s system published through blind peer review without a human in the loop and scored 6.33/10 against a workshop average of 4.87. ARS puts blocking gates at Stage 2.5 and Stage 4.5 and a reviewer that can be calibrated against your own gold set. The trade is throughput for auditability, and the README does not pretend otherwise.

A third influence is PaperOrchestra (Song, Song, Pfister and Yoon, 2026, Google), which inspired v3.3 with Semantic Scholar API verification, an anti-leakage protocol, VLM figure verification and revision-trajectory tracking. ARS implements the trajectory idea through categorical, evidence-anchored criterion trajectories rather than score deltas.

Licence, maintenance and the cost of staying on v3.21.x

The repository's package.json declares CC-BY-NC-4.0, and the README carries a Creative Commons BY-NC 4.0 badge. The GitHub API reports the licence as NOASSERTION, which is a classifier mismatch rather than a second licence, but it is the kind of thing a compliance office will ask about, so check LICENSE and NOTICE.md directly rather than trusting either label. The NC component means non-commercial use; if your work is funded by a commercial sponsor or you intend to sell a derived service, that is a question for your institution, not for this article.

Maintenance is current. The last push was on 2026-08-24, the same day as v3.21.1, and the release cadence is fast: v3.20.1 on 2026-08-16, v3.21.0 on 2026-08-18, v3.21.1 on 2026-08-24. Three releases in eight days. That pace is the upgrade cost. The v3.21.1 notes mention bounded workflow substrates, sealed bakeoffs and transport hardening; v3.21.0 mentions ISO/IEC 42001-spirit transparency, verifiability and a feasibility track; v3.20.1 mentions contract-honesty hardening and bounded evaluation substrates. Each of those touches behaviour that gates output, so a version bump is not purely cosmetic.

The repository ships CHANGELOG.md, GOVERNANCE.md, SECURITY.md, CITATION.cff, CONTRIBUTING.md, MODE_REGISTRY.md and POSITIONING.md, plus evals/, plugin-evals/ and plugin-evals-citation-check/ directories. The presence of MODE_REGISTRY.md is useful if you need to know which modes exist before reading code. If you pin a version for a submission, read the CHANGELOG between your pin and the current release, because the gate classes and thresholds are the parts most likely to move.

Editorial conclusion

Adopt it if you already work inside Claude Code and want the citation and review stages instrumented rather than improvised. Skip it if you need a fully autonomous writing pipeline, since the README states plainly that the tool will not write the paper for you, and skip it if you cannot use a non-commercial licence. Verify first that your Claude Code version supports plugin packaging, that ANTHROPIC_API_KEY is exported, and whether you want the opt-in audit pass, because ARS_CLAIM_AUDIT=1 changes what the formatter will accept.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes