Four Claude Code skills cover research through revision
Academic Research Skills groups its work into deep research, academic paper writing, paper review, and a pipeline that coordinates them. The README lists 8 research modes, 11 paper modes, 6 reviewer modes, and a staged orchestrator. You can ask for a literature review, build an outline, check citations, revise against reviewer comments, or run the longer research-to-paper flow. Its best idea is friction: the pipeline stops for decisions instead of racing toward a finished manuscript.
That scope can look like a paper factory until you read the boundaries. Version 3.22.2 describes a 10-stage orchestrator with mandatory integrity checkpoints at stages 2.5 and 4.5. The project still assigns authorship and responsibility to the researcher. Its simulated reviewers produce author-facing feedback only. They do not make publication decisions, and the software does not submit anything. This is a structured Claude Code working method, with Python checks around selected artifacts, rather than an independent research system.
The 11-second npm install installs no dependencies
Our sandbox cloned commit e79085d into an unprivileged Node 22 container with 3 CPUs, 8 GB of RAM, and no secrets. The 39.2 MB checkout contained 2,776 files and about 275,033 lines of source. Npm finished in 11 seconds, installed 0 packages, and left 60 MB on disk. Npm audit reported 0 known vulnerabilities across every severity level. The package file is private and mainly declares the Pi wrapper, its four skills, and prompts.
There was no npm build target, so we skipped build. There was also no npm test target, so we skipped tests. That distinction matters because the repository itself contains a tests directory and 14 CI workflow files. The lab result does not say those checks failed, and it does not say they passed. It says npm gives a new evaluator no standard command for running them. Reviewers must inspect the project-specific scripts and GitHub workflows if they want to reproduce maintainer checks.
What happened when we ran it
Our run measured repository setup, not academic output. commit e79085d installed 0 packages in 11 seconds, with no build or test script to execute. We did not give Claude a paper, call bibliographic indexes, generate a manuscript, or score a review. The 0-vulnerability npm audit covers the package state npm could see. It cannot assess prompt behavior, Python scripts outside npm's dependency graph, citations returned by a model, or the quality of a drafted argument.
The missing npm targets make the README's own evidence labels especially useful. Release v3.22.2 separates deterministic script checks from prompt-level behavior and repeatedly marks the latter unmeasured. That is unusually candid. It also means adoption should start with one bounded job, such as checking citations in an existing draft, while you retain the inputs and compare every finding against the source. A green dependency audit cannot carry that judgment for you.
Plugin installation keeps controls that other channels lose
The setup guide recommends the Claude Code plugin route. Minimum use needs Claude Code and Anthropic access. Markdown output works without Python. Direct DOCX generation requires Pandoc, and PDF output requires Tectonic plus the listed fonts. A real Python interpreter activates the optional write-scope guard and several commands. Windows support is best-effort, has no Windows CI job, and needs Git Bash for the shell-based hook.
Install choice changes the product. The control availability matrix says the plugin is the reference channel. A skills copy or repository clone does not wire plugin hooks. Cowork runs skills separately without the designed subagent pipeline. A claude.ai Project exposes files as knowledge but executes no routing, hooks, scripts, or orchestration. Claude Science imports only a point-in-time methodology layer. If you want the advertised 10-stage workflow, use the plugin and verify which controls announce themselves.
Its integrity gates cannot prove the study happened
The project's positioning statement draws the most important line on the page. ARS can check citation existence, claim-to-source alignment, reported methods, figures, tables, and package conformance. It cannot establish that procedures were performed, raw data are authentic, or results reproduce. The maintainer gives the blunt consequence: a fabricated study can pass when its false story is internally consistent.
Reviewer confidence is also unfinished. Issue 653 says the reviewer skill has no measured error profile. A planned 12-paper ICLR 2026 calibration set then exposed accept labels through headers, author names, and line-number differences, so issue 828 records a decision to wait for submission-time ICLR 2027 papers before publishing the first profile. That restraint is good science. Until the measurement exists, treat every simulated review as a list of questions to investigate, never a verdict.
CC BY-NC makes it source-available for noncommercial work
GitHub showed 49,900 stars, 31 open issues, 1 open pull request, and a last push on September 29, 2026. Release v3.22.2 arrived four days earlier. Those dates show active maintenance, while the small pull-request count suggests most open work is tracked as issues. The license matters more than the star count: CC BY-NC 4.0 permits sharing and adaptation with attribution for noncommercial use, and the project explicitly says it is not open source.
For a student or noncommercial lab already committed to Claude Code, ARS is worth trying on one paper because it records checkpoints and states where its evidence stops. The 2,776-file checkout is still a large body of instructions to trust. Read the data-flow map, keep cross-model transport off unless you approve the content leaving your machine, and compare every citation with the paper itself. The tool can organize skepticism. It cannot supply the skepticism on your behalf.