CLAUDE.md and Skills: a 3-layer setup ends repeated AI briefings

During a first month with Claude Code, the author was restating the same project facts every session — the stack (Next.js App Router, Supabase, Stripe), the test commands, and the branch-and-PR flow. When an explanation was missed, the AI would try to push straight to main or misread "Kit" as a toolkit and produce odd variable names. His rule of thumb is that anything explained three times belongs in a file, not in the chat.

The proposed split follows how often each file is read rather than what it contains. Facts and terminology go into knowledge markdown files, which are handed over only when needed. Procedures done at least weekly become SKILL.md files under .claude/skills, either project-specific or shared across all projects, with the description field doubling as the trigger condition. Rules that hold every session go into CLAUDE.md at the project root, and the author warns that stuffing long procedures there makes the file fat and less effective, since it is read even on unrelated tasks.

What gives the advice its edge is the pruning discipline. The author lists what to leave out of CLAUDE.md — long procedures, style rules a linter can enforce, one-off instructions, and fast-changing details like current owners and deadlines — and offers a three-question test for where something belongs. He also treats drift as maintenance: rather than adding instructions in chat, he fixes the file, which he argues turns the documentation into an asset that improves with use.