AgentSkillsCN

instincts

在会话过程中捕捉模式与修正建议,将其存储于“健忘记忆”中,并将经过验证的直觉逐步转化为具体技能。

SKILL.md
--- frontmatter
name: instincts
description: Capture patterns and corrections during sessions, store in Forgetful memory, and graduate proven instincts to skills

Instincts - Continuous Learning System

Purpose

Capture patterns and corrections during sessions, store in Forgetful memory, and graduate proven instincts to skills.

What Are Instincts?

Instincts are learned patterns from:

  • User corrections ("No, do it this way...")
  • Repeated preferences observed across sessions
  • Successful approaches to recurring problems
  • Workflow optimizations discovered during work

How Instincts Work

Capture

When a pattern is identified:

  1. Extract the trigger condition ("When X happens...")
  2. Extract the action ("...do Y")
  3. Assign initial confidence based on source:
    • User explicit correction: Medium-High (0.7) - user directly stated preference
    • Observed repeated behavior: Medium-Low (0.55) - inferred from patterns
    • Successful approach: Medium (0.6) - worked but not explicitly requested
  4. Store in Forgetful with instinct tags

Reinforcement

Each time an instinct is used:

  • Positive outcome: Confidence increases by +0.05 (capped at 0.95 to require graduation for full confidence). Upon graduation to a skill, confidence is set to 1.0 (full confidence)
  • User correction: Confidence decreases by -0.15 (floored at 0.1 to preserve learning)
  • Invocation count increments

Graduation

Instincts can evolve to skills when:

  • Confidence > 0.9
  • Invocations > 10
  • No contradictions with existing patterns

Commands

Integration with Forgetful

Instincts are stored as Forgetful memories with special tagging for retrieval and analysis.