New resource note with 6 core methodologies, community best practices, pitfalls, and practical examples. Three zettelkasten notes extract key insights: hook vs prompt reliability, MCP context tradeoffs, and the instinct learning system. Updated existing guides with cross-links.
31 lines
1.1 KiB
Markdown
31 lines
1.1 KiB
Markdown
---
|
||
created: "2026-03-19 12:03"
|
||
type: zettel
|
||
tags: [claude-code, machine-learning, continuous-improvement, agent-evolution]
|
||
source: "https://github.com/affaan-m/everything-claude-code"
|
||
---
|
||
|
||
# 本能学习系统的演化路径
|
||
|
||
ECC 的 Continuous Learning v2.1 实现了一个 AI Agent 自我改进的闭环:
|
||
|
||
```
|
||
观察(Hook捕获) → 模式检测(Haiku模型) → 本能(Instinct) → 技能(Skill)
|
||
```
|
||
|
||
关键设计决策:
|
||
|
||
1. **原子性**: 每个本能只描述一个行为,带信心分数 (0.3-0.9)
|
||
2. **项目隔离**: 用 git remote URL hash 作命名空间,防止跨项目污染
|
||
3. **渐进提升**: 单项目本能 → 多项目验证(2+项目, 信心>=0.8) → 全局技能
|
||
4. **可逆性**: `/evolve` 生成的技能可以回退到本能级别
|
||
|
||
这本质上是一个**强化学习循环** — 用户的接受/拒绝作为奖励信号,信心分数作为 Q-value 近似。与传统 fine-tuning 不同,它在推理时(通过 context injection)而非训练时改变行为,成本低且可控。
|
||
|
||
---
|
||
|
||
## Related
|
||
|
||
- [[Everything Claude Code 方法论与最佳实践]]
|
||
- [[Hook驱动优于提示词驱动]]
|