claesbackman/AI-research-feedback

What it solves

This project provides a suite of specialized AI-driven tools to help academic researchers improve their work before submission. It automates the rigorous process of peer review, reproducibility checks, and grant proposal evaluation, reducing the need for manual pre-submission audits and helping researchers identify weaknesses in their arguments, mathematics, or code alignment.

How it works

The project consists of a collection of "skills" designed for Claude Code. These skills are installed as Markdown files in the Claude Code environment and are triggered via slash commands (e.g., /review-paper). Depending on the skill, the system employs different multi-agent architectures—such as an 8-agent parallel review for full papers or a 6-agent panel for grant proposals—to analyze LaTeX files, analysis code (Stata, R, Python), and supporting documentation. It can simulate specific journal personas (e.g., AER, QJE) to tailor the feedback to specific editorial standards.

Who it’s for

Academic researchers, particularly those in economics and finance, who use LaTeX for writing and need high-quality, critical feedback on papers, pre-analysis plans, and grant proposals.

Highlights

• Multi-Agent Peer Review: Simulates a full editorial board with specialized agents focusing on grammar, internal consistency, mathematics, and opposing "advocate" and "skeptic" perspectives.
• Paper-Code Alignment: Specifically checks if the empirical claims in a LaTeX paper are accurately reflected in the accompanying analysis code.
• Grant & PAP Review: Dedicated workflows for evaluating grant proposals (NSF, NIH, ERC) and Pre-Analysis Plans (PAPs).
• Code Change Explainer: Generates an offline HTML page with a quiz to help coauthors understand the empirical consequences of code changes in a research pipeline.