400 lines in a single claude.md.

That was my Claude Code setup 6 months ago. Claude behaved differently every session. The moment a team member picked up the same project, they had to re-explain everything from scratch.

We blamed the model. The problem was the file.

The fix is one structural decision made once at the start of any project: separate what Claude knows from how it behaves from where it is in a workflow right now.

Three folders handle most of it:

rules/ – behavioral instructions, committed to git, applied across every task

context/ – domain knowledge, updated as the project evolves, never mixed with instructions

memory/ – runtime state only, gitignored, reset between sessions

We now create a Claude Project, attach the GitHub repo in project settings, and every team member links their own project to the same repo. One committed folder structure becomes the team's shared context. One day to scaffold. Day one outputs match day 90, regardless of who is running it.

One step most people skip: after any update to the repo, sync it in your Claude Project settings.

Claude does not pull changes automatically. Miss this once and you will not know you are running on stale context until the outputs start drifting.

This week's FutureBrief covers the full 8-folder structure, 4 pitfalls with fixes, and a GitHub template with everything ready to clone.