Metaskill

通过脚本引导 AI 代理在出错后做表层、原则和习惯三层分析，在任务前检索过往经验，并在成功后记录有效模式；支持多种 LLM 提供商，也可回退到本地启发式模式。

fahrulalwan

热度 58.08

最近更新 2026-05-11 17:27:43

分类：知识依赖：代码执行 / 读写文件 / 内部调用大模型 / 第三方 API应用场景：笔记整理 / 工作流 / 调试排错 / 研究

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add metaskill

Metadata

name: metaskill
description: Teaches AI agents how to learn better by enforcing deep correction, transfer learning, and proactive pattern recognition. Use when an error occurs and needs deep analysis (not surface patch), before starting a complex task to check past analogies, or after success to capture winning patterns. NOT for routine file reads or simple one-off commands.

Metaskill

3 Core Components

Deep Self-Correction (deep-correct.sh) — 3-level breakdown on errors:
- Surface: What specifically failed
- Principle: The underlying rule/constraint violated
- Habit: Concrete behavioral change to prevent recurrence
Transfer Learning (transfer-check.sh) — Before a task, search past learnings for analogous patterns. Maps domains (e.g., "auth" → "security") to prevent siloed learning.
Proactive Pattern Recognition (success-capture.sh) — Log what worked and why, building a repository of successful patterns.

Usage

# When an error occurs
bash skills/metaskill/scripts/deep-correct.sh "description of the error"

# Before starting a complex task
bash skills/metaskill/scripts/transfer-check.sh "description of the new task"

# After successful execution
bash skills/metaskill/scripts/success-capture.sh "what worked" "why it worked"

# Monthly health eval
bash skills/metaskill/scripts/eval.sh --save

Configuration (LLM Provider)

Metaskill uses two provider tiers — fast (extraction) and deep (transfer/eval). Edit config.yaml to match your setup:

# config.yaml
providers:
  fast: anthropic   # change to: openai | ollama | gemini
  deep: anthropic

Provider	Env Var	Notes
`anthropic`	`ANTHROPIC_API_KEY`	Default
`openai`	`OPENAI_API_KEY`
`ollama`	(none needed)	Local, free
`gemini`	`GOOGLE_API_KEY`

Ollama example (fully local, no API key):

providers:
  fast: ollama
  deep: ollama
models:
  ollama:
    fast: llama3.2
    deep: llama3.1:70b

If no provider is available, metaskill falls back to manual/heuristic mode (still works, but less precise extraction).

Integration with Self-Improving-Agent

Writes to skills/self-improving-agent/.learnings/ if present, otherwise falls back to its own .learnings/ directory. No extra setup needed.

AGENTS.md Wiring (Mandatory)

Add to pre-task checklist:

Run transfer-check.sh before any major task
Run deep-correct.sh immediately after any error (not just LEARNINGS.md append)
Run success-capture.sh after complex task completes successfully