A Better Way to Give Claude Code Context

I just finished my first deep dive with Claude Code, a Vision proof of concept that extracts handwritten data off clear plastic bags—filled plastic bags, that is …. Claude Code and I started Friday, and by Monday we had a robust proof of concept, implemented in Python CLI form, including parallelized LLM calls and detailed documentation. My working time was maybe an hour per day, and the Anthropic API bill for Claude’s model usage was about $80. I’d rate the ROI on that $80 investment extremely high ….

I knew enough to spend time giving Claude plenty of context. I was at the AI Engineer World’s Fair in June, had seen Anthropic’s Boris Cherny speak. So I had Cherny’s context tips, and created a number of markdown files, including CLAUDE.md at the project root, along with a well-populated /docs folder.

I had noted Scott Werner’s post Nobody Knows How To Build With AI Yet when it first came out a week or two back, and finally got around to reading it this morning. Werner levels us up on how to give Claude Code context, adding some well-thought-out separation and structure:

His full GitHub Gist here. As Werner caveats, though:

I have no idea what I’ll build next week. What process I’ll use. What documents I’ll create or abandon. Whether any of this will still make sense … My four-document system isn’t a recommendation. It’s just one data point in the collective experiment we’re all running. A fossil from last week’s development process. Already outdated. Already nostalgic.

Things move fast in AI time, but for now, I’ll apply Werner’s approach on my next Claude Code project.