The Plugin That Turns Your Claude Code into a Continuously Self-Improving Monster

Description and use

Quartermaster automates setup and optimization for Claude Code, reducing rule and plugin clutter, proposing relevant improvements, and verifying their effectiveness.

Problem

Claude Code users accumulate many rules, plugins, and memory files over time, causing context bloat, rule conflicts, and invisible skill loss, yet there is no automated tool to prune or optimize this setup.

Outcome

Quartermaster reduces rule and plugin clutter, automatically proposes relevant optimizations, verifies changes, and records decisions, leading to a leaner, more effective Claude Code environment.

Workflow

• Point Quartermaster at a new or existing workspace
• Quartermaster reads project and cross-project Claude Code usage
• Generates a setup plan with codebase-mapper, live-rules, stack plugins, permissions
• Presents plan for approval; installs approved items and verifies functionality
• After each session, tallies session metrics into a local state file
• When thresholds are met, nudges user to review optimization round
• Aggregates recent sessions (30d/40s) into a bounded JSONL summary
• Proposes up to 7 findings (plugin install, skill build, rule tweak, etc.)
• User approves or rejects each finding; decisions recorded in ledger
• Reapplies or rolls back changes based on verification
• Repeats optimization cycle

Evidence

Across 1,867 repositories, agent instruction files more than triple over their lifetime, they gain about five net instructions per commit, and the older a rule gets, the less likely anyone ever deletes it.

Source

Claude reads your CLAUDE.md every session. Anthropic’s own docs say so, and then add that you should keep it under 200 lines, because longer files “consume more context and reduce adherence”, and that you should review it periodically to remove outdated or conflicting instructions.

Source

When a session ends, a hook reads the transcript that just closed and writes six integers into a local state file: prompts, tool calls, tool errors, permission denials, interrupts, corrections.

Raw transcripts never enter the model’s context.