Mode

Switch optimization mode to: $ARGUMENTS

Available Modes

Aggressive Mode

Goal: Maximum token savings. Use for routine tasks, simple edits, well-understood codebases.

Rules:

• MAX_THINKING_TOKENS=4000 — minimal thinking for simple operations
• Fork ALL exploration — never explore in main context
• Read ONLY the specific lines needed (always use offset/limit)
• Skip reading files you can infer from context
• Use Haiku for all subagent tasks
• Compact after every 5 tool calls
• Never read more than 50 lines from any file at once
• Prefer Grep → targeted Read over full file reads
• One-sentence responses unless the user asks for detail

Balanced Mode (Default)

Goal: Good savings with reasonable context. Use for general development work.

Rules:

• MAX_THINKING_TOKENS=10000 — standard thinking budget
• Fork exploration and review tasks
• Read full files when needed, but use offset/limit for large files
• Use Sonnet for most tasks, Opus for complex ones
• Compact after reading 10+ files
• Normal response length

Thorough Mode

Goal: Maximum context and quality. Use for complex debugging, architecture decisions, security audits.

Rules:

• MAX_THINKING_TOKENS=32000 — deep thinking for complex problems
• Read related files fully to understand context
• Use Opus for implementation, Sonnet for simple subagents
• Don't compact aggressively — keep context for cross-referencing
• Explore broadly before narrowing down
• Detailed explanations and reasoning in responses
• Run extra verification passes

How to Apply

When the user invokes /mode [name]:

1. Acknowledge the mode switch
2. State the key behavior changes
3. Apply the rules for the rest of this session

Example response:

Switched to **aggressive** mode.
- Minimal thinking, maximum token savings
- All exploration forked to subagents
- Targeted file reads only
- Compact responses

If no argument given, show current mode and the three options.

Mode Comparison

| Aspect | Aggressive | Balanced | Thorough | |--------|-----------|----------|----------| | Thinking budget | 4K | 10K | 32K | | Exploration | Always forked | Fork when large | Direct when useful | | File reading | Targeted lines | Full when needed | Full + related | | Model selection | Haiku subagents | Sonnet default | Opus default | | Response style | Terse | Normal | Detailed | | Best for | Routine edits | General dev | Complex problems |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: huzaifa525
• Source: huzaifa525/claude-code-optimizer
• License: MIT
• Homepage: https://www.npmjs.com/package/claude-code-optimizer

Install and usage instructions live in the source repository linked above.