My Claude Code Was Wasting 63K Tokens Per Session. Here's How I Fixed It.

Background

Here’s a fun fact I discovered last week: every time I opened Claude Code, 153KB of markdown files were silently injected into my system prompt — before I even typed a single word. Blogging plugins. Note-taking plugins. Fourteen workflow skills I rarely used outside my main projects.

That’s 63K tokens gone before the conversation even starts. Over 20 sessions a day? Over a million tokens burned on nothing.

The fix took six lines of JSON. Here’s the full breakdown.

1. The Problem

Before touching anything, I needed to understand where tokens were going.

find ~/.claude/plugins/cache -name "SKILL.md" -exec wc -c {} \;
claude --debug --debug-file /tmp/debug.txt --print "hello"

The debug output revealed the full picture:

That’s ~153KB of markdown instructions loaded into the system prompt before I even typed a message, plus 5,430 chars of hook context. At ~2.5 bytes per token for mixed English/code content, that’s roughly ~63K tokens of overhead per new session.

The principle was simple: load only what you need, only where you need it. Claude Code supports three mechanisms:

1. Global disable — set plugins to false in ~/.claude/settings.json
2. Project-level settings — .claude/settings.json in project dirs re-enables plugins locally
3. Runtime flags — --plugin-dir loads a plugin for one session

2. The Fix

Global settings (~/.claude/settings.json) — disable 3 plugins, lower effort, stratify models:

- "superpowers@superpowers-marketplace": true,
- "obsidian@obsidian-skills": true,
- "frontend-design@claude-plugins-official": true,
+ "superpowers@superpowers-marketplace": false,
+ "obsidian@obsidian-skills": false,
+ "frontend-design@claude-plugins-official": false,

- "CLAUDE_CODE_EFFORT_LEVEL": "max",
+ "CLAUDE_CODE_EFFORT_LEVEL": "high",

- "ANTHROPIC_SMALL_FAST_MODEL": "deepseek-v4-pro[1m]",
+ "ANTHROPIC_SMALL_FAST_MODEL": "deepseek-v4-flash",

Project-level overrides — re-enable only where needed:

# ~/Work/trade/exchange-rs/.claude/settings.json
{ "enabledPlugins": { "superpowers@superpowers-marketplace": true } }

# ~/Work/dex/.claude/settings.json
{ "enabledPlugins": { "superpowers@superpowers-marketplace": true } }

# ~/Documents/LocalKnowledge/AI-Brain/AI-Brain/.claude/settings.json
{ "enabledPlugins": { "obsidian@obsidian-skills": true } }

That’s it. Three files, six edits. Rust-analyzer-LSP stays global since it’s a language server with no skill overhead.

3. Results

To verify the optimization actually worked, I ran controlled tests using --debug:

# Test in a neutral directory (no project settings)
cd /tmp && claude --debug --debug-file /tmp/after-debug.txt --print "hello"

Comparing debug outputs before and after:

The key debug log line:

# Before:  getSkills returning: ... 18 plugin skills ...
# After:   getSkills returning: ... 0 plugin skills ...

The optimization is transparent for active projects — enter the directory and Claude Code automatically picks up the local settings.json, loading exactly the plugins needed. For one-off tasks: claude --plugin-dir <path>.

4. Takeaways

• Profile first: find ... -name "SKILL.md" revealed 153KB hidden overhead. Can’t optimize what you can’t measure.
• Lazy-load plugins: Project-level settings.json scopes plugins to specific directories. Same pattern as direnv / asdf.
• Tune effort level: max dramatically increases thinking tokens. high is enough for most tasks — save max for hard problems.
• Stratify models: Not every operation needs the flagship model. Use flash variants for background tasks.
• Numbers compound: ~63K tokens saved per session × 20 sessions = 1.26M tokens per day.