Use a coding agent to run the data analysis for a research paper

Kevin Yang, a researcher who builds data pipelines and figures for papers, used Claude Code for the analysis on one of his papers. He estimates it wrote over 98% of the code. He works with it in a terminal next to Cursor for review and Jupyter for checking output. His method is to have it write an overall plan in Markdown, then a detailed plan per part, such as cleaning, each research question and the final figures, then one script per step. He reviews each script and its results before moving on and uses git to roll back bad changes. It still makes frequent mistakes, so validation stays his job. He found other coding agents weaker for this.