module 13 tokens performance

Trim overhead

System Text-to-Speech Ready
Slide: 0:00 / 0:00
Slide 1 of 0Interactive Deck

Full Lesson Reference

Your baseline token overhead (what loads before you type) is the most leveraged number to optimise. Cut 20K off your baseline = 20K less per message × every message forever. This lesson is the audit + trim routine.

The 5 sources of overhead

  1. System prompt (~10K, fix ed)
  2. CLAUDE.md files (target 1-3K)
  3. Skill metadata (16K hard cap, target 10-12K)
  4. MCP schemas (target 5-10K, varies by deferred loading)
  5. Startup memory (~1-3K, from /startup)

Total healthy baseline: 25-35K. Bloated baseline: 50K+. Over 60K = audit immediately.

Audit your baseline

Start a fresh session + run /startup. Then

Before I type anything else - show me exactly what's loaded in this session and the approximate token cost of each:

  • CLAUDE.md files (global + project)
  • Skills (each with metadata size)
  • MCPs (each with connection status + approximate schema cost)
  • Memory loaded from /startup

Give me a total and flag anything over-sized.

Claude breaks it down. You see where the bloat is.

Trim CLAUDE.md

Covered in detail in Module 02. Tell Claude in plain English (no special skill needed):

Audit my global CLAUDE.md. For each line: does it apply to every project? Would removing it cause mistakes? Flag for deletion, movement, or keep. Target: under 25 lines.

Same audit for project CLAUDE.md. Target 30-60 lines.

Trim skills

Module 09 Lesson 4. Quick version

Audit my skills. Report usage vs 16K budget. Flag skills I haven't used in 60+ days. List descriptions over 400 characters.

Remove unused, tighten descriptions, move project-specific skills to project folders.

Trim MCPs

The biggest overhead source for most users. Audit with

List all connected MCPs in this session. For each, show the approximate schema size + whether I've invoked it in the last week. Flag any I should disconnect.

MCPs that add up silently

  • One-click connects you forgot about
  • MCPs from shared skills that auto-connected
  • enableAllProjectMcpServers: true in some settings (loads EVERY MCP, overrides per-project config)
  • Old experiments you tried once and never disconnected

Check for the "load all" flag

Check if enableAllProjectMcpServers is true in any of my project settings. Change to false and list only the servers I actually use.

Real-world example

A power user ran this audit

  • Before: 280 MCP tool definitions loading every message = ~50K overhead
  • After: ~50 tools = ~12K overhead
  • Per session: 500K total tokens → 300K total tokens (40% cheaper)

One audit. Half the cost for the same output.

Per-project vs global

What should be global (applies to every session) vs project-specific (applies only when working in that project)?

Global (load every session)

  • Your identity + locale (you don't change)
  • Universal skills (wrapup, startup, data-integrity)
  • Cross-platform MCPs you use daily (memory Supabase, GitHub CLI)

Project-specific (load only in that project)

  • Client-specific MCPs (Google Ads account X, Meta account Y)
  • Client-specific skills (brand enforcers, custom workflows)
  • Project-level CLAUDE.md rules Move anything client-specific OUT of global. Your global baseline stays low.

The trim routine

Monthly 10-minute audit

  1. Check baseline token usage in a fresh session
  2. If > 40K, audit skills (target trim 5-10K)
  3. If still high, audit MCPs (biggest potential win)
  4. Audit CLAUDE.md for creep
  5. Test with a fresh session - should feel snappier

Power-user tips

  • Keep a minimum-viable stack - only what you use this week. Install more when you need them.
  • Disable before disconnecting - toggle MCPs off first to see if you miss them. If not, disconnect permanently.
  • Use Supermetrics for tail platforms - one MCP covers 150+ sources; beats installing 10 separate small MCPs
  • Archive skills you might want back - move to ~/.claude/skills-archive/ instead of deleting

Action items

☐ Run the baseline audit in a fresh session

☐ If over 40K baseline, start with MCPs (biggest win)

☐ Audit skills against 16K budget

☐ Audit CLAUDE.md files for creep

☐ Monthly trim routine on the calendar

Next lesson: The 50% rule + session discipline.

Exercises

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