What they did
Koreeda does life-science research on an older Intel MacBook with no GPU. Each paper lives in its own GitHub repo, bundled as git submodules into one workspace where he starts Claude Code, with per-paper CLAUDE.md files and skills, and generated wrapper skills at the root. All data sits in Snowflake: raw public data, tables shaped per research question with comments that explain what a row means, and odd file formats in S3 behind an external stage. Claude Code queries it with the Snowflake CLI and an MCP server over Cortex Agents, picks its own language and environment, writes analysis results back to Snowflake, and pushes code, figures and Markdown manuscripts to GitHub. A separate skill collects prior work as original texts and stops on evidence rules. Hooks speak aloud when Claude Code needs permission or finishes, and he steers it from his phone through Remote Control.
How it went
He says he no longer opens RStudio or Docker and rarely opens an editor except to edit manuscripts. Snowflake costs averaged about $23 a month from May to October 2026, about $6 of it storage for 270 GB, with the busiest warehouse month at $20. Analyses can be rerun from Snowflake in a fresh worktree with no local data, and he checks that the 20 main result files match. He gives no figures on time saved or on errors the agent made.
Worth knowing
Table comments do real work here: he says Claude Code reads them before writing SQL and stopped misreading what values mean. He also records settled decisions in CLAUDE.md so new sessions do not reopen them. The author sells a book on Snowflake cost optimization and links it in the post.