ClaudeGLP1 — Claude Get Lean Protocol 1

A 5 minute Claude Code skill that audits your Claude environment, estimates how much context window you're burning at session open, and guides you through an interactive cleanup — with a full backup and one-command restore.

---

What it does

If you've been trying out agent tools like /GSD and others, your Claude is probably wasting tons of tokens. Each Claude Code session loads a surprising amount of content before you type anything: CLAUDE.md files, skills, hooks, memory files, MCP server configs, and system-injected tool lists. ClaudeGLP1 measures all of it, flags what's redundant or misplaced, and helps you trim it down — it takes about 5 minutes.

The 8 phases:

Phase	Name	What happens
0	Welcome	Friendly orientation — explains what's coming and waits for your go-ahead
1	Discovery	Scans CLAUDE.md hierarchy, skills, hooks, plugins, and memory files
2	Token Report	Estimates tokens consumed at session open, per component
3	Recommendations	Flags quick wins and structural improvements with token savings
4	Backup	Snapshots your entire current setup before touching anything
5	Cleanup	Walks through each recommendation — you approve every change
6	New Hierarchy	Shows your refreshed CLAUDE.md structure as a plain-text org chart
7	Before/After	Side-by-side token comparison: before vs. after
8	Restore	On-demand rollback to any previous backup

---

Requirements

• Claude Code installed and authenticated
• macOS or Linux (Windows path adjustments may be needed)

---

Installation

From GitHub

# Clone the repo
git clone https://github.com/<username>/ClaudeGLP1.git
cd ClaudeGLP1

# Copy the skill into your Claude commands directory
cp ClaudeGLP1.md ~/.claude/commands/ClaudeGLP1.md

Manual

Download ClaudeGLP1.md and copy it directly:

cp /path/to/ClaudeGLP1.md ~/.claude/commands/ClaudeGLP1.md

That's it. No dependencies, no configuration required. Claude Code will pick up the new skill automatically — no restart needed.

Verify installation

Start a Claude Code session and run:

/ClaudeGLP1

If Phase 0 appears with the welcome message, you're set.

---

Updating

cd ClaudeGLP1
git pull
cp ClaudeGLP1.md ~/.claude/commands/ClaudeGLP1.md

---

Usage

Start a full audit:

/ClaudeGLP1

Restore a previous backup:

/ClaudeGLP1 --restore glp1_20260413_143022

List available backups (tell Claude to list them):

/ClaudeGLP1 --restore

---

What gets backed up

Before any changes are made, ClaudeGLP1 saves:

• All CLAUDE.md files in your hierarchy (global, vault, project levels)
• All files in ~/.claude/commands/
• ~/.claude/settings.json and any project-level settings files
• Memory files (recent-memory.md, long-term-memory.md, project-memory.md)

Backups are stored at ~/.claude/backups/glp1_<timestamp>/ with a MANIFEST.md that records every file's original path and the pre-cleanup token baseline.

---

What it looks for

Quick wins:

• Unused or redundant skills in ~/.claude/commands/
• Placeholder content in CLAUDE.md files (stub tables, empty sections)
• Hooks with "verbose": true that inject output into every conversation turn
• Memory files that are empty or contain only boilerplate

Structural issues:

• Rules in a CLAUDE.md that belong at a different level in the hierarchy
• Content duplicated across global / vault / project levels
• MCP servers or plugins that are configured but not actively used

---

CLAUDE.md hierarchy

Claude Code supports three levels of configuration, loaded from outermost to innermost:

~/.claude/CLAUDE.md              ← global (all projects)
└── ~/your-vault/CLAUDE.md       ← vault or workspace level
    └── ~/your-vault/project/CLAUDE.md   ← per-project

ClaudeGLP1 helps you put the right rules at the right level so nothing is duplicated and context overhead is minimized.

---

Token estimation

ClaudeGLP1 uses ~4 characters per token as a conversion approximation. System-injected content (deferred tools list, skills reminder, date injection) is estimated from observed sizes and noted as approximate. Actual usage may vary slightly by model and session.

---

Notes

• ClaudeGLP1 never makes changes without your explicit confirmation per item.
• Backups are never deleted automatically — they accumulate in ~/.claude/backups/ and can be restored at any time.
• The skill has no external dependencies and works entirely with your local filesystem.
• Tested with Claude Code CLI on macOS. Should work on Linux; Windows paths may need adjustment.