I published my Claude Code starter prompt as a downloadable file: STARTPROMPT.md

It sets Claude up as technical lead for any project, runs a structured discovery session, then creates five .md files as external memory — CLAUDE.md, PROGRESS.md, DECISIONS.md, TECHNICAL.md, and HANDOVER.md. Each has explicit read/write rules so context doesn't rot across sessions.

The longer explanation is in the blog post. The short version: don't let your AI carry project state in conversation history. Put it in files with a protocol for when to read and write them.