Claude Code — Token-Saving Config

A curated collection of Claude Code configuration files (global and per-project) focused on reducing token consumption without sacrificing quality.

Structure

.
├── global/               # User-level config (~/.claude/)
│   ├── CLAUDE.md         # Global system prompt directives
│   ├── RTK.md            # RTK hook environment directives
│   └── settings.json     # Global settings (plugins, hooks, env vars)
│
└── per-project/          # Drop into project root or .claude/
    └── code-review-graph/
        ├── CLAUDE.md     # MCP tool directives for code-review-graph
        ├── settings.json # Project hooks (PostToolUse, SessionStart)
        └── .mcp.json     # MCP server definition

Key Token-Saving Strategies

Strategy	How	Savings
Minimal output rules	CLAUDE.md — no filler, minimal code blocks	High
Autocompact at 50% context	CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50	Medium
Capped thinking tokens	MAX_THINKING_TOKENS=10000	Medium
RTK bash output stripping	PreToolUse hook rewrites verbose CLI output	High
effortLevel: "low"	Reduces reasoning effort on simple tasks	Medium
Knowledge graph (MCP)	Replaces Grep/Glob/Read with graph queries	High
claude-mem summaries	Compresses session history across conversations	Medium
caveman response compression	Rewrites Claude output in terse fragment style	High

Global Plugins

Plugin	Repo	Purpose
RTK	rtk-ai/rtk	Hook that strips boilerplate from CLI output
claude-mem	thedotmack/claude-mem	Session memory & token economics stats
claude-hud	jarrodwatts/claude-hud	Status line with token usage info
caveman	JuliusBrussee/caveman	Compresses Claude output to terse fragments (~65% token savings)

Per-Project Setups

Project	Repo	Strategy
code-review-graph	tirth8205/code-review-graph	Knowledge graph MCP replaces file scanning

Global Settings Highlights

• effortLevel: "low" — Claude uses less reasoning effort by default.
• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: "50" — Context is compacted at 50% full instead of the default, preventing large context blowups.
• MAX_THINKING_TOKENS: "10000" — Hard cap on extended thinking token budget.
• RTK PreToolUse hook — Every Bash command is piped through rtk-rewrite.sh which strips headers, banners, and verbose logs before the output reaches Claude.

claude-mem Configuration

Settings used with the claude-mem plugin:

Setting	Value
Observations	10
Sessions	10
Display Count	0
Display Filed	Narrative
Token economics (read cost, work investment, savings)	All enabled
AI Provider	Gemini API — gemini-2.5-flash-lite
Rate limiting	Active
Include last summary	Active
Include last message	Disabled

Using Gemini as the AI provider for claude-mem keeps memory processing off Claude API credits entirely.

Caveman Configuration

Settings used with the caveman plugin:

Setting	Value
Intensity	Full (default) — articles and filler removed, fragments used
Activation	/caveman or "less tokens please"
Deactivation	"stop caveman" or "normal mode"
caveman-compress	/caveman:compress CLAUDE.md — rewrites memory files with ~46% fewer input tokens

Intensity levels: lite (grammar-safe) → full (default) → ultra (telegraphic) → wenyan (classical Chinese)

caveman-compress preserves code, URLs, and paths — only prose is compressed. Original is saved as CLAUDE.md.original.