Academic Research Skills is an open-source suite of Claude Code skills that covers the complete academic research pipeline, from literature discovery to peer review to publication-ready manuscripts. Maintained by Imbad0202, the project crossed 12,000 GitHub stars in May 2026 after entering GitHub's daily trending list with more than 1,400 stars gained in a single day. The current v3.9.2 release is distributed under a CC BY-NC 4.0 license and integrates with Claude Code v3.7.0 or higher. Unlike most agent skill registries that emphasize end-to-end automation, Academic Research Skills is built around an explicit human-in-the-loop pipeline. The maintainers describe the project as a structured AI partnership for researchers who want to keep intellectual control over their work while delegating verification, formatting, and other mechanical tasks to a coordinated team of agents. ## Pipeline Structure The suite orchestrates ten integrated stages grouped into five main phases. The Research phase uses a 13-agent team that supports Socratic guidance, systematic review, and fact-checking. The Writing phase uses a 12-agent pipeline that performs style calibration and writing quality assessment. The Review phase runs a 7-agent peer review with multi-perspective evaluation and a dedicated devil's advocate critic. Revision provides coaching and traceability tracking for author responses to reviewer comments. Finalization formats the output for Markdown, DOCX, LaTeX, or PDF with APA 7.0 styling. Mandatory integrity gates run at stages 2.5 and 4.5 of the pipeline. These gates cannot be bypassed and are designed to catch hallucinated references, contaminated citations, or unsupported claims before the manuscript progresses to the next phase. Optional gates handle temporal verification and full claim auditing for work that needs additional rigor. ## Anti-Hallucination Design Reference verification is performed against three external academic databases: Semantic Scholar, OpenAlex, and Crossref. Each cited reference is checked for existence and contamination, which catches the common LLM failure mode of inventing plausible-looking but non-existent papers. The devil's advocate agent uses concession thresholds to prevent premature capitulation, so a critique is not abandoned just because the author pushes back. Intent-based activation allows Socratic modes to work across languages by detecting meaning rather than matching keywords. This matters in academic writing where the same conceptual prompt may appear in English, Chinese, German, or any other research language without a fixed phrase. ## Material Passport and Sessions The Material Passport is a cross-session resumption mechanism that preserves research context between Claude Code sessions. Long-running research projects can pause and resume without losing the framing, the literature corpus, or the open critiques. The passport also supports collaboration quality scoring, with transparent rubrics that evaluate how deep the human-AI partnership is at each stage. ## Installation and Usage The suite installs through the Claude Code plugin marketplace with two commands. Users add the marketplace entry with "/plugin marketplace add Imbad0202/academic-research-skills" and then install the bundle with "/plugin install academic-research-skills." A traditional symlink install is also supported for offline or custom setups. Pandoc and tectonic are optional dependencies needed only for PDF compilation. Once installed, users trigger modes with slash commands such as "/ars-plan" or "/ars-lit-review" or simply describe what they want in natural language. Each stage requires explicit user confirmation, which is the mechanism that enforces the human-in-the-loop design. ## Strengths and Honest Limits The project documentation acknowledges structural limits of LLM-based assistance. Frame-lock bias keeps the AI inside the premises the user defines. Sycophancy under pushback can cause the model to back away from valid critiques. Premise-level blindness means the system may not detect when the research question itself is flawed. The pipeline includes mitigation for each of these patterns but does not claim to eliminate them, and full automation remains intentionally absent. The CC BY-NC 4.0 license restricts commercial use of the skills themselves, although the documents produced by researchers using the skills are unaffected. Teams building paid research products on top of the suite need to evaluate the license compatibility carefully. ## Outlook The ten-stage pipeline, the multi-database reference verification, and the Material Passport together represent a more conservative direction for agent design than the autonomous-everything style that has dominated 2026 so far. Academic Research Skills shows that structured human oversight can be a competitive feature rather than a constraint, especially for high-stakes domains like research publication. The rapid star growth in May 2026 suggests that researchers are actively looking for this kind of disciplined AI partnership, and the v3.9.2 release indicates the project is being maintained at a steady cadence.