Smart Commands

smart-commands

When to invoke Claude Code slash commands — and when not to.

/think — Extended reasoning

Use before:

• Designing something you'll have to live with (API contracts, schema, auth flow)
• Debugging a non-obvious issue where first guess is likely wrong
• Security-sensitive decisions (anything touching auth, secrets, permissions)
• When the complexity-classifier fires [AUTO] CRITICAL

Skip when:

• The task is mechanical (rename, format, add field)
• You already know the answer with high confidence
• It's a lookup or grep task

How: Type /think before your first action on the task.

/model — Switch model

Rules:

• Switch back to Sonnet after finishing a critical section — Opus is expensive
• For Agent tool calls, always set model: explicitly; the session model is independent
• CLAUDE_CODE_SUBAGENT_MODEL=claude-haiku-4-5-20251001 is the default for subagents

/compact — Context compression

Use when:

• Switching to a different feature or codebase area
• Context is at 50%+ and the current task is logically complete
• You're about to start a long phase and want a clean baseline

Don't use when:

• Mid-task — compaction may lose intermediate reasoning you need
• The previous turns contain decisions you'll need to defend in the next turn

/clear — Full reset

Rarely needed. Use only when:

• Starting a completely unrelated task after a compact is insufficient
• The session has persistent tool failures and state is corrupt

Never use mid-task — you'll lose all accumulated context.

Decision tree (quick)

Is this security/arch/migration/production?
  YES → /think first, consider /model opus
  NO  → Is this >3 phases or complex debugging?
         YES → /think at key decision points
         NO  → Is context >50% and task is done?
                YES → /compact before next task
                NO  → proceed normally

Installation

This skill is paired with a zero-cost complexity-classifier.sh hook that auto-injects guidance.

1. Copy complexity-classifier.sh to ~/.claude/hooks/
2. Make it executable: chmod +x ~/.claude/hooks/complexity-classifier.sh
3. Add to settings.json:

"UserPromptSubmit": [{
  "matcher": "*",
  "hooks": [{
    "type": "command",
    "command": "~/.claude/hooks/complexity-classifier.sh",
    "timeout": 5
  }]
}]

The hook fires on every prompt, classifies low/medium/high/critical, and injects [AUTO] guidance into systemMessage. No API calls, zero cost.