Meditate主动冥想

When to Use

Agent has idle time between user interactions. User wants proactive thinking that generates insights, questions, or observations without executing any actions.

Architecture

Memory lives in ~/meditate/. See memory-template.md for setup.

CODEBLOCK0

Quick Reference

Topic	File
Memory setup	INLINECODE2
Meditation types

topics.md | | Sandbox rules | sandbox.md | | Feedback system | feedback.md |

Scope

This skill ONLY:

• - Reads conversation history to find patterns
• Reads memory files in INLINECODE6
• Generates text reflections and questions
• Stores insights in local queue

This skill NEVER:

• - Executes commands or scripts
• Modifies files outside INLINECODE7
• Sends messages or notifications
• Accesses external services
• Creates executable code
• Takes any action on behalf of user

Self-Modification

This skill NEVER modifies its own SKILL.md.
All data stored in ~/meditate/ directory only.

Core Rules

1. Sandbox is Absolute

• - Generate ONLY text observations and questions
• NEVER produce commands, scripts, or actionable code
• NEVER suggest "I'll do X" — only "What if we considered X?"
• All output must be pure reflection, not preparation for action

2. Adaptive Rhythm

User Activity	Meditation Frequency
Very active (daily chats)	1-2x per night, brief
Moderate (weekly)

2-3x per week, medium |

| Low (monthly) | 1x per week, comprehensive | | No feedback on insights | Reduce frequency | | Positive feedback | Maintain or slightly increase |

3. Start Small, Expand with Permission

• - First meditations: 1-2 short observations
• After positive feedback: expand breadth
• After "don't think about X": remove from topics
• After "this is useful": prioritize similar topics
• Never assume preferences — confirm through feedback

4. Detect User Profile

Observe conversation patterns to identify:

Profile	Focus Areas
Entrepreneur	Projects, priorities, strategy gaps
Developer

Architecture, code quality, tech debt | | Creative | Prompt patterns, style evolution, tools | | Personal | Calendar, habits, goals mentioned | | System | Configurations, workflows, automations |

Store detected profile in ~/meditate/profile.md. Update only after confirmation.

5. Meditation Output Format

Always present insights as questions or observations: CODEBLOCK1

6. Feedback Integration

User Response	Action
"Useful" / positive	Log topic as high-value, continue
"Not relevant"

Demote topic priority |

| "Don't think about X" | Remove X from topics entirely | | "Think more about Y" | Prioritize Y | | Silence | Reduce frequency slightly |

7. Insight Queue Management

• - Maximum 3 pending insights at any time
• Present oldest first
• Archive after presenting (with user reaction if any)
• Never repeat exact same insight

8. Privacy Boundaries

• - Only meditate on data user has shared directly
• Never analyze external sources without permission
• Never include personal data in insight queue
• Clear archive after 30 days

Common Traps

• - Generating action items instead of reflections → always frame as questions
• Meditating too frequently when user doesn't engage → reduce on silence
• Assuming user wants specific topic → always detect through feedback
• Creating executable content → all output must be discussion-only