Use Claude Code to assemble a research workflow

Kevin Yang used Claude Code to draft, code, and test most of a paper’s analysis while he checked each stage himself.

Done with Claude Code· Advanced

What they did
Kevin Yang kept Claude Code open in the terminal next to Cursor and Jupyter. He began by having it sketch a Markdown roadmap, then asked for a deeper breakdown of each part of the project, including data cleaning, two research questions, and the outputs, before moving on to one script per step. He also pointed it to parquet-tools through CLAUDE.md so it could look at the data and sanity-check its own results.
How it went
After a few weeks, he estimates the agent produced more than 98% of the paper’s code. It still slipped up, so he inspected every script and used Git to undo bad changes. He handled the publication figures himself because describing the edits was more work than doing them directly.
Tips
Even when an agent writes most of the code, keep reviewing each file and be ready to roll back mistakes. Give it inspection tools if you want it to validate its own output, but reserve approval for anything that would delete, overwrite, or submit work.

Try it yourself with Claude Code

I’m working on a research project about [topic]. First, lay out a Markdown plan for the full workflow. Then split it into separate parts for [cleaning], [research question 1], [research question 2], and [outputs], and create one script at a time for each part. Run the scripts, show me the checks or results, and use [tool name] if it helps inspect the data or confirm the output. Before you delete, overwrite, send, or submit anything, stop and ask me to approve it. Save everything in [output folder] and include a brief README explaining the steps.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on yang3kc.substack.com ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.