Run an AI research workflow with human checkpoints
Experience a workflow that handles drafting, reference checking, and peer review in one go. This tool is a collection of skills for Claude Code that guides you from defining a research topic to finalizing a manuscript, with human confirmation points at each stage. The AI handles sentence generation and formatting, allowing researchers to focus on core logic and methodology selection.
A structure where humans decide and AI organizes
The tool avoids full automation and adopts a 'human-in-the-loop' approach, requiring user confirmation at each stage. While the AI drafts text or searches literature, users make key judgments such as defining research questions, selecting methodologies, and interpreting data. This ensures that humans ultimately filter out hallucinated citations or logical errors that the AI might generate. The tool does not hide the fact that it is being used; instead, it focuses on learning the user's writing style to reduce mechanical tone.
Step-by-step features from research to publication
The tool provides a five-stage pipeline: research, writing, review, revision, and finalization. In the research stage, it uses Socratic dialogue to establish the paper structure and performs systematic literature reviews. In the writing stage, 12 agents collaborate to draft the text, reflecting the user's past writing style. The review stage generates feedback from various perspectives, including journal suitability, methodology, and logical counterarguments. The final stage verifies citation accuracy and data consistency to ensure manuscript integrity.
Installation and execution environment
You need Claude Code version v3.7.0 or higher and an ANTHROPIC_API_KEY. You can install it with two lines of commands from the plugin marketplace. If you want DOCX or PDF output, you must additionally install Pandoc and tectonic, but these are not required for Markdown output only. Python is needed only for some optional commands and the writing scope guard, as core features operate on a prompt basis. After installation, you can start with the /ars-plan command to begin a conversation about planning the paper structure.
Limitations of verification and precautions
The tool checks for the existence of citations, the alignment between claims and sources, and the consistency of reported methodologies, but it does not verify whether experiments were actually performed, the authenticity of original data, or the reproducibility of results. Consistently reported false information can pass this verification, so final responsibility remains with the author. Additionally, the current live review feature is in a 'NOT_CALIBRATED' state, and the application of a full calibration profile has not yet been connected.
