What files load when Claude Code starts — and in what order
Progressive disclosure: load only what you need, when you need it
When you run claude, it doesn't load everything at once. It follows a three-phase pattern:
Phase 1 — User Level: Your personal settings and preferences from ~/.claude/
Phase 2 — Project Level: Infrastructure-wide context from ~/projects/CLAUDE.md
Phase 3 — On Demand: Detailed reference files loaded only when a skill or agent needs them
This keeps the initial context small and fast. Claude gets relevant details only when the task calls for them.
~/.claude/ — personal preferences, always first
~/projects/CLAUDE.md — infrastructure-wide context
Context has a cost: larger context = slower responses + higher token usage. Detailed network topology, security policies, and logging configs are only relevant for specific tasks. Loading them on-demand keeps boot fast and focused.
network-config, troubleshooting, service-deploymentkeycloak-setup, secrets-security, traefik-setuplogging-observability skill only when setting up monitoring.The key principle: Agents and skills are discovered at boot (names + triggers only), but their full content loads only when invoked. This is progressive disclosure — Claude knows what tools exist without paying the cost of loading all of them upfront.
Claude Code CLI Boot Sequence • ai-servicers.com