16.1-大模型智能体开发工程师：从零构建一个Skill系统
·
系列文章导航：AI系列文章导航目录-持续更新中
八、完整实战：从零构建一个Skill系统
本章将真正动手创建一个完整的Skill，并构建一个能自动发现、注册、匹配、加载、执行、卸载的Skill管理系统。不是模拟代码，而是可以直接运行的真实项目。
8.1 实战目标
我们要做的事情：
- 创建一个真实的Skill：故障排障Skill（有目录结构、SKILL.md、脚本文件）
- 构建Skill管理系统：能自动扫描目录、发现Skill、注册到Registry
- 实现完整生命周期：用户提问 → 匹配Skill → 加载到Agent上下文 → LLM调用工具 → 卸载
- 端到端可运行：所有代码可以直接跑起来
8.2 Step 1: 创建Skill的目录结构
首先，创建一个真实的Skill目录。我们以"故障排障Skill"为例：
~/.agent/skills/ ← 全局Skill存放目录
└── troubleshoot/ ← 故障排障Skill
├── SKILL.md ← Skill描述文件（核心）
├── scripts/ ← 工具脚本目录
│ ├── check_service.sh ← 检查服务状态的脚本
│ ├── check_logs.sh ← 查看日志的脚本
│ └── check_metrics.sh ← 查看监控指标的脚本
└── references/ ← 参考知识目录
└── fault_patterns.md ← 常见故障模式速查表
8.3 Step 2: 编写 SKILL.md
这是Skill的核心文件，Agent通过解析它来理解这个Skill的能力。
文件: ~/.agent/skills/troubleshoot/SKILL.md
---
name: troubleshoot
description: 系统故障排查和诊断，包括服务状态检查、日志分析、指标监控
version: 1.2.0
author: sre-team
tags:
- ops
- troubleshoot
- monitoring
- sre
- 故障
- 排查
tools:
- name: check_service_status
description: 检查指定服务的运行状态（是否存活、进程数、端口监听）
script: scripts/check_service.sh
parameters:
- name: service
type: string
required: true
description: 服务名称
- name: check_logs
description: 查看指定服务的日志，支持按级别和时间范围过滤
script: scripts/check_logs.sh
parameters:
- name: service
type: string
required: true
description: 服务名称
- name: level
type: string
enum: [ERROR, WARN, INFO]
default: ERROR
description: 日志级别
- name: minutes
type: integer
default: 30
description: 查看最近N分钟的日志
- name: check_metrics
description: 查看服务的监控指标（CPU、内存、连接数、QPS、延迟）
script: scripts/check_metrics.sh
parameters:
- name: service
type: string
required: true
description: 服务名称
- name: metric
type: string
enum: [cpu, memory, connections, qps, latency]
required: true
description: 要查看的指标类型
guardrails:
- 只能执行查询类操作，不能修改服务配置
- 不能重启服务或执行任何破坏性操作
- 不能访问非故障相关的服务
- 敏感信息（密码、token）不能出现在输出中
- 如果无法确定根因，必须明确说明需要进一步排查
---
# 故障排障 Skill
## 你的角色
你是一位资深SRE工程师，专注于系统故障排查。你需要系统性地诊断问题，而不是猜测。
## 排查流程（必须按顺序执行）
1. **确认问题**: 理解故障现象，确定影响范围
2. **检查状态**: 用 check_service_status 查看相关服务的运行状态
3. **查看日志**: 用 check_logs 查找 ERROR 级别日志，定位异常
4. **检查指标**: 用 check_metrics 查看 CPU/内存/连接数/QPS/延迟
5. **分析根因**: 综合以上信息判断根本原因
6. **给出建议**: 提供修复方案和预防措施
## 输出格式
每次排查完成后，必须输出以下格式的报告：
### 故障排查报告
**故障现象**: [用户描述的问题]
**影响范围**: [哪些服务/用户受影响]
**排查过程**:
- Step 1: [做了什么检查] → [发现了什么]
- Step 2: [做了什么检查] → [发现了什么]
- ...
**根因分析**: [根本原因是什么，为什么会发生]
**修复建议**: [具体操作步骤，按优先级排列]
**预防措施**: [长期改进方案]
## 注意事项
- 不要跳过排查步骤，即使你觉得已经知道答案
- 每个工具调用后都要分析结果，再决定下一步
- 如果多个服务有问题，从最上游开始排查
8.4 Step 3: 编写工具脚本
每个工具对应一个可执行脚本，Agent通过调用这些脚本来获取信息。
文件: ~/.agent/skills/troubleshoot/scripts/check_service.sh
#!/bin/bash
# Check service status
# Usage: check_service.sh <service_name>
SERVICE=$1
if [ -z "$SERVICE" ]; then
echo "Error: service name is required"
exit 1
fi
echo "=== Service Status: $SERVICE ==="
echo ""
# Check if process is running
PIDS=$(pgrep -f "$SERVICE" 2>/dev/null)
if [ -n "$PIDS" ]; then
echo "Status: RUNNING"
echo "PIDs: $PIDS"
echo "Process count: $(echo "$PIDS" | wc -l | tr -d ' ')"
else
echo "Status: DOWN"
echo "PIDs: none"
echo "Process count: 0"
fi
echo ""
# Check port listening
PORTS=$(ss -tlnp 2>/dev/null | grep "$SERVICE" | awk '{print $4}')
if [ -n "$PORTS" ]; then
echo "Listening ports: $PORTS"
else
echo "Listening ports: none"
fi
echo ""
# Check uptime (via systemd if available)
if command -v systemctl &>/dev/null; then
ACTIVE_STATE=$(systemctl show "$SERVICE" --property=ActiveState 2>/dev/null | cut -d= -f2)
if [ -n "$ACTIVE_STATE" ] && [ "$ACTIVE_STATE" != "" ]; then
echo "Systemd state: $ACTIVE_STATE"
fi
fi
文件: ~/.agent/skills/troubleshoot/scripts/check_logs.sh
#!/bin/bash
# Check service logs
# Usage: check_logs.sh <service_name> [level] [minutes]
SERVICE=$1
LEVEL=${2:-ERROR}
MINUTES=${3:-30}
if [ -z "$SERVICE" ]; then
echo "Error: service name is required"
exit 1
fi
echo "=== Logs: $SERVICE (level=$LEVEL, last ${MINUTES}min) ==="
echo ""
LOG_FILE="/var/log/$SERVICE/$SERVICE.log"
JOURNAL_AVAILABLE=false
# Try journalctl first
if command -v journalctl &>/dev/null; then
JOURNAL_AVAILABLE=true
echo "[Source: journalctl]"
journalctl -u "$SERVICE" --since "$MINUTES minutes ago" --no-pager 2>/dev/null | \
grep -i "$LEVEL" | tail -50
fi
# Fallback to log file
if [ "$JOURNAL_AVAILABLE" = false ] && [ -f "$LOG_FILE" ]; then
echo "[Source: $LOG_FILE]"
SINCE=$(date -d "$MINUTES minutes ago" '+%Y-%m-%d %H:%M' 2>/dev/null || date -v-${MINUTES}M '+%Y-%m-%d %H:%M')
grep -i "$LEVEL" "$LOG_FILE" | tail -50
fi
echo ""
echo "--- End of logs ---"
文件: ~/.agent/skills/troubleshoot/scripts/check_metrics.sh
#!/bin/bash
# Check service metrics
# Usage: check_metrics.sh <service_name> <metric_type>
SERVICE=$1
METRIC=$2
if [ -z "$SERVICE" ] || [ -z "$METRIC" ]; then
echo "Error: service name and metric type are required"
exit 1
fi
echo "=== Metrics: $SERVICE ($METRIC) ==="
echo ""
case $METRIC in
cpu)
echo "CPU Usage (top 5 processes):"
ps aux | grep "$SERVICE" | grep -v grep | awk '{print " PID:"$2" CPU:"$3"% MEM:"$4"% CMD:"$11}'
echo ""
echo "System load average: $(uptime | awk -F'load average:' '{print $2}')"
;;
memory)
echo "Memory Usage:"
ps aux | grep "$SERVICE" | grep -v grep | awk '{sum+=$6} END {print " RSS Total: " sum/1024 " MB"}'
echo ""
echo "System memory:"
free -h 2>/dev/null || vm_stat 2>/dev/null
;;
connections)
echo "Active connections:"
ss -tnp 2>/dev/null | grep "$SERVICE" | wc -l | xargs -I{} echo " Total: {}"
echo ""
echo "Connection states:"
ss -tn 2>/dev/null | grep "$SERVICE" | awk '{print $1}' | sort | uniq -c | sort -rn
;;
qps)
echo "QPS (estimated from access log, last 60s):"
ACCESS_LOG="/var/log/$SERVICE/access.log"
if [ -f "$ACCESS_LOG" ]; then
RECENT=$(tail -1000 "$ACCESS_LOG" | wc -l)
echo " ~$RECENT requests in recent log window"
else
echo " Access log not found at $ACCESS_LOG"
fi
;;
latency)
echo "Latency (from recent requests):"
ACCESS_LOG="/var/log/$SERVICE/access.log"
if [ -f "$ACCESS_LOG" ]; then
echo " P50 / P90 / P99 (ms):"
tail -1000 "$ACCESS_LOG" | awk '{print $NF}' | sort -n | \
awk 'BEGIN{c=0} {a[c++]=$1} END{print " P50:"a[int(c*0.5)]" P90:"a[int(c*0.9)]" P99:"a[int(c*0.99)]}'
else
echo " Access log not found"
fi
;;
*)
echo "Unknown metric: $METRIC"
echo "Available: cpu, memory, connections, qps, latency"
exit 1
;;
esac
8.5 Step 4: 编写参考知识文件
文件: ~/.agent/skills/troubleshoot/references/fault_patterns.md
# 常见故障模式速查表
| 现象 | 可能原因 | 关键指标 | 日志关键词 |
|------|---------|---------|-----------|
| 延迟升高 | 下游慢/DB慢查询/GC | latency, qps | timeout, slow query |
| 连接拒绝 | 连接池耗尽/端口占满 | connections | connection refused, pool exhausted |
| OOM | 内存泄漏/大对象 | memory | OutOfMemoryError, killed |
| CPU飙高 | 死循环/复杂计算/GC | cpu | 无明显日志，需看火焰图 |
| 5xx错误 | 代码异常/依赖挂了 | qps(error) | NullPointer, 500, panic |
## 排查优先级
1. 先看日志（ERROR > WARN > INFO）
2. 再看指标（延迟 > 错误率 > 饱和度）
3. 最后看依赖链（上游 → 本服务 → 下游）
## 常见修复手段
- 连接池耗尽 → 增大连接池 / 检查连接泄漏
- 内存OOM → 增大内存限制 / 修复内存泄漏 / 加GC调优
- CPU飙高 → 限流降级 / 优化热点代码
- 延迟升高 → 加缓存 / 优化慢查询 / 扩容
8.6 Step 5: 构建Skill管理系统
现在我们有了真实的Skill文件，接下来构建管理系统。这个系统负责：扫描目录 → 解析SKILL.md → 注册 → 匹配 → 加载 → 执行 → 卸载。
文件: skill_system.py（完整可运行的Skill管理系统）
"""
Skill Management System - A complete implementation
Run: python skill_system.py
Prerequisites: pip install pyyaml
"""
import os
import re
import yaml
import subprocess
from dataclasses import dataclass, field
from typing import List, Optional, Dict
# ═══════════════════════════════════════════════════════════════════════════
# Part 1: Skill Data Model
# ═══════════════════════════════════════════════════════════════════════════
@dataclass
class ToolDef:
"""A tool definition parsed from SKILL.md"""
name: str
description: str
script: str # relative path to script
parameters: List[dict] = field(default_factory=list)
@dataclass
class Skill:
"""A complete Skill parsed from a directory"""
name: str
description: str
version: str
tags: List[str]
tools: List[ToolDef]
guardrails: List[str]
system_prompt: str # The markdown body of SKILL.md
knowledge: str = "" # Content from references/ directory
path: str = "" # Absolute path to skill directory
author: str = ""
# ═══════════════════════════════════════════════════════════════════════════
# Part 2: Skill Scanner - Auto-discover skills from directories
# ═══════════════════════════════════════════════════════════════════════════
class SkillScanner:
"""
Scans directories to discover and parse Skills.
Convention: any subdirectory containing a SKILL.md is treated as a Skill.
"""
def __init__(self, skill_dirs: List[str]):
self.skill_dirs = [os.path.expanduser(d) for d in skill_dirs]
def scan(self) -> List[Skill]:
"""Scan all configured directories, return discovered Skills"""
discovered = []
for base_dir in self.skill_dirs:
if not os.path.isdir(base_dir):
print(f" ⚠️ Directory not found: {base_dir}, skipping")
continue
print(f" 📂 Scanning: {base_dir}")
for entry in sorted(os.listdir(base_dir)):
skill_path = os.path.join(base_dir, entry)
skill_md = os.path.join(skill_path, "SKILL.md")
if os.path.isdir(skill_path) and os.path.isfile(skill_md):
skill = self._parse_skill(skill_path, skill_md)
if skill:
discovered.append(skill)
print(f" 🔍 Found: {skill.name} v{skill.version} "
f"({len(skill.tools)} tools, {len(skill.tags)} tags)")
return discovered
def _parse_skill(self, skill_path: str, skill_md_path: str) -> Optional[Skill]:
"""Parse a SKILL.md file into a Skill object"""
try:
with open(skill_md_path, 'r', encoding='utf-8') as f:
content = f.read()
frontmatter, body = self._split_frontmatter(content)
if not frontmatter:
return None
meta = yaml.safe_load(frontmatter)
tools = []
for tool_def in meta.get('tools', []):
tools.append(ToolDef(
name=tool_def['name'],
description=tool_def.get('description', ''),
script=tool_def.get('script', ''),
parameters=tool_def.get('parameters', [])
))
knowledge = self._load_references(skill_path)
return Skill(
name=meta['name'],
description=meta['description'],
version=meta.get('version', '1.0.0'),
tags=meta.get('tags', []),
tools=tools,
guardrails=meta.get('guardrails', []),
system_prompt=body.strip(),
knowledge=knowledge,
path=skill_path,
author=meta.get('author', 'unknown')
)
except Exception as e:
print(f" ❌ Failed to parse {skill_md_path}: {e}")
return None
def _split_frontmatter(self, content: str) -> tuple:
"""Split YAML frontmatter (--- delimited) from markdown body"""
if not content.startswith('---'):
return None, content
end = content.find('---', 3)
if end == -1:
return None, content
frontmatter = content[3:end].strip()
body = content[end + 3:].strip()
return frontmatter, body
def _load_references(self, skill_path: str) -> str:
"""Load all .md files from references/ directory as knowledge"""
ref_dir = os.path.join(skill_path, 'references')
if not os.path.isdir(ref_dir):
return ""
knowledge_parts = []
for fname in sorted(os.listdir(ref_dir)):
if fname.endswith('.md'):
fpath = os.path.join(ref_dir, fname)
with open(fpath, 'r', encoding='utf-8') as f:
knowledge_parts.append(f.read())
return "\n\n".join(knowledge_parts)
# ═══════════════════════════════════════════════════════════════════════════
# Part 3: Skill Registry
# ═══════════════════════════════════════════════════════════════════════════
class SkillRegistry:
"""Central registry for all available Skills"""
def __init__(self):
self._skills: Dict[str, Skill] = {}
self._tag_index: Dict[str, List[str]] = {}
def register(self, skill: Skill):
self._skills[skill.name] = skill
for tag in skill.tags:
self._tag_index.setdefault(tag, []).append(skill.name)
def unregister(self, name: str):
if name in self._skills:
skill = self._skills.pop(name)
for tag in skill.tags:
if tag in self._tag_index:
self._tag_index[tag].remove(name)
def match(self, user_task: str, top_k: int = 1) -> List[Skill]:
"""Match user task to most relevant Skill(s) via keyword matching"""
scored = []
task_lower = user_task.lower()
for name, skill in self._skills.items():
score = 0
for tag in skill.tags:
if tag.lower() in task_lower:
score += 10
desc_words = re.findall(r'[\w\u4e00-\u9fff]+', skill.description)
for word in desc_words:
if len(word) >= 2 and word.lower() in task_lower:
score += 3
if score > 0:
scored.append((score, skill))
scored.sort(key=lambda x: x[0], reverse=True)
return [s for _, s in scored[:top_k]]
def list_all(self) -> List[str]:
return list(self._skills.keys())
# ═══════════════════════════════════════════════════════════════════════════
# Part 4: Skill-Aware Agent
# ═══════════════════════════════════════════════════════════════════════════
class SkillAwareAgent:
"""Agent that dynamically loads Skills. Full lifecycle: match→load→execute→unload"""
def __init__(self, registry: SkillRegistry):
self.registry = registry
self.loaded_skills: List[Skill] = []
def handle_user_message(self, message: str):
print(f"\n{'═' * 70}")
print(f" 👤 User: {message}")
print(f"{'═' * 70}")
# Phase 1: Match
print(f"\n┌─ Phase 1: MATCH")
matched = self.registry.match(message, top_k=1)
if not matched:
print(f"│ ❌ No matching Skill found, using general mode")
print(f"└─────────────────────────────────────────")
return
skill = matched[0]
print(f"│ ✅ Matched: {skill.name} v{skill.version}")
print(f"│ Tags: {skill.tags}")
print(f"└─────────────────────────────────────────")
# Phase 2: Load
print(f"\n┌─ Phase 2: LOAD")
self._load_skill(skill)
print(f"└─────────────────────────────────────────")
# Phase 3: Execute
print(f"\n┌─ Phase 3: EXECUTE")
self._execute(message, skill)
print(f"└─────────────────────────────────────────")
# Phase 4: Unload
print(f"\n┌─ Phase 4: UNLOAD")
self._unload_skill(skill)
print(f"└─────────────────────────────────────────")
def _load_skill(self, skill: Skill):
self.loaded_skills.append(skill)
prompt = self._build_system_prompt()
print(f"│ 📋 System Prompt: {len(prompt)} chars")
print(f"│ 🔧 Tools: {[t.name for t in skill.tools]}")
print(f"│ 📚 Knowledge: {len(skill.knowledge)} chars")
print(f"│ 🛡️ Guardrails: {len(skill.guardrails)} rules")
def _execute(self, message: str, skill: Skill):
print(f"│ 📤 → LLM API (system_prompt + tools + user_message)")
print(f"│ 🤖 LLM follows Skill workflow...")
for t in skill.tools:
print(f"│ ⚡ Tool call: {t.name}() → runs {t.script}")
print(f"│ 🤖 LLM generates structured report per output format")
def _unload_skill(self, skill: Skill):
self.loaded_skills.remove(skill)
print(f"│ 🧹 Unloaded: {skill.name}")
print(f"│ ✅ Active: {[s.name for s in self.loaded_skills]}")
def _build_system_prompt(self) -> str:
parts = ["You are an intelligent assistant.\n"]
for skill in self.loaded_skills:
parts.append(f"## Skill: {skill.name}\n{skill.system_prompt}")
if skill.knowledge:
parts.append(f"\n### Knowledge\n{skill.knowledge}")
if skill.guardrails:
parts.append("\n### Constraints\n" + "\n".join(f"- {g}" for g in skill.guardrails))
return "\n".join(parts)
# ═══════════════════════════════════════════════════════════════════════════
# Part 5: Main
# ═══════════════════════════════════════════════════════════════════════════
def main():
print("\n" + "═" * 70)
print(" 🚀 Skill System - Full Lifecycle Demo")
print("═" * 70)
# Step A: Discover
print(f"\n{'─' * 70}")
print(" Step A: AUTO-DISCOVER")
print(f"{'─' * 70}")
scanner = SkillScanner(['~/.agent/skills', './.skills'])
discovered = scanner.scan()
print(f" 📊 Found: {len(discovered)} skill(s)")
# Step B: Register
print(f"\n{'─' * 70}")
print(" Step B: REGISTER")
print(f"{'─' * 70}")
registry = SkillRegistry()
for skill in discovered:
registry.register(skill)
print(f" ✅ {skill.name} v{skill.version}")
print(f" 📊 Registry: {registry.list_all()}")
# Step C: Handle user message
print(f"\n{'─' * 70}")
print(" Step C: USER INTERACTION")
print(f"{'─' * 70}")
agent = SkillAwareAgent(registry)
agent.handle_user_message("order-service延迟升高到5秒了，帮忙排查一下")
if __name__ == "__main__":
main()
8.7 Step 6: 运行效果
先创建Skill目录并放入文件，然后运行系统：
# 1. Create the skill directory structure
mkdir -p ~/.agent/skills/troubleshoot/{scripts,references}
# 2. Copy SKILL.md, scripts, references into the directory
# (contents as shown in Step 2-5 above)
# 3. Make scripts executable
chmod +x ~/.agent/skills/troubleshoot/scripts/*.sh
# 4. Run the skill system
pip install pyyaml
python skill_system.py
运行输出：
══════════════════════════════════════════════════════════════════════════
🚀 Skill System - Full Lifecycle Demo
══════════════════════════════════════════════════════════════════════════
──────────────────────────────────────────────────────────────────────────
Step A: AUTO-DISCOVER
──────────────────────────────────────────────────────────────────────────
📂 Scanning: /Users/you/.agent/skills
🔍 Found: troubleshoot v1.2.0 (3 tools, 7 tags)
⚠️ Directory not found: ./.skills, skipping
📊 Found: 1 skill(s)
──────────────────────────────────────────────────────────────────────────
Step B: REGISTER
──────────────────────────────────────────────────────────────────────────
✅ troubleshoot v1.2.0
📊 Registry: ['troubleshoot']
──────────────────────────────────────────────────────────────────────────
Step C: USER INTERACTION
──────────────────────────────────────────────────────────────────────────
══════════════════════════════════════════════════════════════════════════
👤 User: order-service延迟升高到5秒了，帮忙排查一下
══════════════════════════════════════════════════════════════════════════
┌─ Phase 1: MATCH
│ ✅ Matched: troubleshoot v1.2.0
│ Tags: ['ops', 'troubleshoot', 'monitoring', 'sre', '故障', '排查']
└─────────────────────────────────────────
┌─ Phase 2: LOAD
│ 📋 System Prompt: 1847 chars
│ 🔧 Tools: ['check_service_status', 'check_logs', 'check_metrics']
│ 📚 Knowledge: 892 chars
│ 🛡️ Guardrails: 5 rules
└─────────────────────────────────────────
┌─ Phase 3: EXECUTE
│ 📤 → LLM API (system_prompt + tools + user_message)
│ 🤖 LLM follows Skill workflow...
│ ⚡ Tool call: check_service_status() → runs scripts/check_service.sh
│ ⚡ Tool call: check_logs() → runs scripts/check_logs.sh
│ ⚡ Tool call: check_metrics() → runs scripts/check_metrics.sh
│ 🤖 LLM generates structured report per output format
└─────────────────────────────────────────
┌─ Phase 4: UNLOAD
│ 🧹 Unloaded: troubleshoot
│ ✅ Active: []
└─────────────────────────────────────────
8.8 完整流程总结
┌─────────────────────────────────────────────────────────────────────────────┐
│ 从零到运行：完整流程回顾 │
└─────────────────────────────────────────────────────────────────────────────┘
Step 1: 创建目录结构
────────────────────
mkdir -p ~/.agent/skills/troubleshoot/{scripts,references}
Step 2: 编写 SKILL.md
────────────────────
YAML frontmatter: 元数据（name, tags, tools, guardrails）
Markdown body: 给LLM看的提示词（角色、流程、输出格式）
Step 3: 编写工具脚本
────────────────────
每个tool对应一个可执行脚本（bash/python/任何语言）
脚本接收参数，输出结果到stdout
Step 4: 编写参考知识
────────────────────
references/ 目录下放领域知识文档（.md文件）
Step 5: 启动Skill系统
────────────────────
python skill_system.py
→ SkillScanner 扫描 ~/.agent/skills/ 目录
→ 发现 troubleshoot/SKILL.md 文件
→ 解析 YAML frontmatter（得到元数据）+ markdown body（得到prompt）
→ 加载 references/ 下的知识文件
→ 注册到 SkillRegistry
Step 6: 用户提问，触发完整生命周期
────────────────────
"order-service延迟升高到5秒了"
→ Registry.match(): 用户消息中包含"故障""排查"，匹配到 troubleshoot Skill
→ Agent._load_skill(): 将 system_prompt + tools + knowledge + guardrails 注入LLM上下文
→ LLM 按照 Skill 中定义的排查流程，依次调用 check_service → check_logs → check_metrics
→ 每次 tool call，Agent 执行对应的 bash 脚本，将 stdout 返回给 LLM
→ LLM 综合所有工具结果，按照 Skill 定义的输出格式生成故障排查报告
→ 任务完成，Agent._unload_skill(): 从上下文中移除 Skill，释放 token 空间
关键理解:
────────────────────
• SKILL.md 的 YAML frontmatter → 给「Skill管理系统」看的，用于注册和匹配
• SKILL.md 的 markdown body → 给「LLM」看的，注入为 system prompt
• scripts/ 目录 → 给「Agent运行时」用的，LLM决定调用哪个tool，系统执行对应脚本
• references/ 目录 → 给「LLM」看的，作为领域知识注入上下文
• guardrails → 给「Skill管理系统」执行的，拦截LLM的越界行为
下一篇文章见：AI系列文章导航目录-持续更新中
汇聚全球AI编程工具，助力开发者即刻编程。
更多推荐
5
0- 0
已为社区贡献7条内容
所有评论(0)