What loads at start

Every time you type claude, 4 things load before you type your first prompt. Knowing what loads tells you why sessions feel fast or slow, and where to look when something's wrong.

What loads, in order

1. CLAUDE.md files

Your global CLAUDE.md (~/.claude/CLAUDE.md) loads first - your identity, preferences, safety rules. Then the project-level CLAUDE.md in the folder you started Claude in, if one exists.

These are pure context. Every instruction in them applies to this session.

2. Memory (if /startup is run)

When you run /startup, Claude connects to your Supabase memory layer and reads:

• The project record for this folder (context, pending items, rules)
• The most recent session summary for this project
• Global rules that apply across all projects

Now Claude knows what you did last time and where to pick up.

3. Skills

Claude scans ~/.claude/skills/ and registers every skill as a slash command. Each skill's description (not full content) loads into context. Type / to see what's available. Skills stay cheap at session start - only the short descriptions load. The full skill content only loads when you actually invoke one.

4. Connectors (MCPs + CLIs)

MCPs configur ed for this project activate. The project's .mcp.json file says which MCPs to connect to - so a project configur ed for Google Ads + Meta connects to those, not every platform you have installed.

CLIs (like gh for GitHub, supabase for the database) don't load anything at session start - they only use resources when Claude actually runs a command.

What "loaded" means

Loaded = in Claude's context window = counts against your token budget.

Heavier loads mean

• Slightly higher cost per message
• Less context window for actual work
• Slower response on the first message

Lean loads mean fast, cheap sessions. Module 13 (Tokens) covers how to keep loads lean.

Check what's loaded

Tell Claude anytime

Show me everything currently loaded in this session. CLAUDE.md files, skills, MCPs, and approximate token cost.

Claude lists everything with line counts and token estimates. Useful when sessions feel slow - you'll spot an MCP or skill you forgot to disable.

What DOESN'T load automatically

These only come in when you explicitly use them

• Files - you tell Claude to read a file, it reads
• MCP tool schemas - deferred loading, only when you call a tool
• Skill bodies - only when you invoke a slash command
• Database queries - only when Claude asks Supabase for something

Session start is a lean baseline. What you DO in the session decides the rest.

Action items

☐ Memorise the 4 things that load at session start: CLAUDE.md + memory + skills + connectors

☐ Know that skill and MCP content are deferred - they don't load fully until called

☐ Practice asking Claude: "show me what's loaded right now"

Next lesson: The context window.

Exercises

1. Review the concepts covered in this lesson: What loads at start.
2. Write down your key takeaway from this lesson.
3. Practice running any commands or prompts mentioned above inside your terminal.