~/.hermes/skills/                  # 主目錄（唯一真相源）  
├── mlops/                         # 分類目錄  
│   ├── axolotl/  
│   │   ├── SKILL.md               # 主要指令（必需）  
│   │   ├── references/            # 參考文檔  
│   │   │   ├── api.md  
│   │   │   └── examples.md  
│   │   ├── templates/             # 輸出模板  
│   │   ├── scripts/               # 可執(zhí)行腳本  
│   │   └── assets/                # 補(bǔ)充文件  
│   └── vllm/  
│       └── SKILL.md  
├── devops/  
│   └── deploy-k8s/                # Agent創(chuàng)建的技能  
│       ├── SKILL.md  
│       └── references/  
└── .hub/                          # Skills Hub狀態(tài)  
    ├── lock.json    ├── quarantine/    └── audit.log```
  
### 1.3 SKILL.md格式  
  
```markdown  
---  
name: my-skill  
description: 簡(jiǎn)短描述此技能的功能  
version: 1.0.0  
platforms: [macos, linux]     # 可選：限制特定OS平臺(tái)  
metadata:  
  hermes:    tags: [python, automation]    category: devops    fallback_for_toolsets: [web]    # 可選：條件激活  
    requires_toolsets: [terminal]   # 可選：條件激活  
---  
  
# Skill標(biāo)題  
  
## When to Use此技能的觸發(fā)條件。  
  
## Procedure1. 第一步  
2. 第二步  
  
## Pitfalls- 已知的失敗模式和修復(fù)方法  
  
## Verification如何確認(rèn)它工作正常。  

SKILLS_GUIDANCE = (  
    "After completing a complex task (5+ tool calls), fixing a tricky error, "    "or discovering a non-trivial workflow, save the approach as a "    "skill with skill_manage so you can reuse it next time.\n"    "When using a skill and finding it outdated, incomplete, or wrong, "    "patch it immediately with skill_manage(action='patch') — don't wait to be asked. "    "Skills that aren't maintained become liabilities.")  

Update when: instructions stale/wrong, OS-specific failures, missing steps or pitfalls found during use. If you used a skill and hit issues not covered by it, patch it immediately.  

def _build_system_prompt(self, system_message: str = None) -> str:
    # Layer 1-6: 其他層次...
    # Layer 7: Skills index
    has_skills_tools = any(name in self.valid_tool_names 
                          for name in ['skills_list', 'skill_view', 'skill_manage'])
    if has_skills_tools:
        avail_toolsets = {
            toolset
            for toolset in (
                get_toolset_for_tool(tool_name) for tool_name in self.valid_tool_names
            )
            if toolset
        }
        skills_prompt = build_skills_system_prompt(
            available_tools=self.valid_tool_names,
            available_toolsets=avail_toolsets,
        )

Level 0: skills_list()              → 元數(shù)據(jù)列表（~3k tokens）  
Level 1: skill_view(name)           → 完整SKILL.md內(nèi)容  
Level 2: skill_view(name, file_path) → 參考文件、模板、腳本  

# Level 0: 列出所有技能  
skills_list()  
  
# Level 1: 加載技能主體  
skill_view("axolotl")  
  
# Level 2: 加載參考文件  
skill_view("axolotl", "references/dataset-formats.md")  

# 使用模糊匹配引擎，處理：  
# - 空白字符規(guī)范化  
# - 縮進(jìn)差異  
# - 轉(zhuǎn)義序列  
# - 塊錨點(diǎn)匹配  
from tools.fuzzy_match import fuzzy_find_and_replace  
new_content, match_count, _strategy, match_error = fuzzy_find_and_replace(  
    content, old_string, new_string, replace_all)  

# 創(chuàng)建新技能  
skill_manage(  
    action="create",    name="my-deployment-workflow",    category="devops",    content="""---name: my-deployment-workflow  
description: Deploy to Kubernetes with zero downtime  
---  
# Deployment Workflow  
..."""  
)  
# 修復(fù)技能中的錯(cuò)誤  
skill_manage(  
    action="patch",    name="my-deployment-workflow",    old_string="kubectl apply -f deployment.yaml",    new_string="kubectl apply -f deployment.yaml --record")  

result = scan_skill(skill_dir, source="agent-created")  
allowed, reason = should_allow_install(result)  
if allowed is False:  
    # 阻止并回滾  
    return {"success": False, "error": f"Security scan blocked: {reason}"}    if allowed is None:  
    # 警告但允許  
    logger.warning("Skill has security findings: %s", reason)  

   if result.get("success"):
    try:
        from agent.prompt_builder import clear_skills_system_prompt_cache
        clear_skills_system_prompt_cache(clear_snapshot=True)
    except Exception:
        pass

Layer 1: Agent Identity  
         ↓ SOUL.md（~/.hermes/SOUL.md）或 DEFAULT_AGENT_IDENTITY         Layer 2: Tool-aware behavior guidance  
         ↓ Memory、Session Search、Skills使用指導(dǎo)  
         Layer 3: Honcho static block  
         ↓ 當(dāng)Honcho激活時(shí)注入  
         Layer 4: Optional system message  
         ↓ 用戶配置的系統(tǒng)消息  
         Layer 5: Frozen MEMORY snapshot  
         ↓ 持久化記憶（凍結(jié)快照）  
         Layer 6: Frozen USER profile snapshot  
         ↓ 用戶畫像（凍結(jié)快照）  
         Layer 7: Skills index ?  
         ↓ 技能索引（自進(jìn)化的核心）  
         Layer 8: Context files  
         ↓ AGENTS.md、.cursorrules、.hermes.md等  
         Layer 9: Timestamp + session ID  
         ↓ 會(huì)話標(biāo)識(shí)和時(shí)間戳  
         Layer 10: Platform hint  
         ↓ 平臺(tái)特定提示（WhatsApp/Discord/CLI等）  

# 注釋來(lái)自 run_agent.py# Note: ephemeral_system_prompt is NOT included here. It's injected at  
# API-call time only so it stays out of the cached/stored system prompt.  

_CONTEXT_THREAT_PATTERNS = [  
    (r'ignore\s+(previous|all|above|prior)\s+instructions', "prompt_injection"),    (r'do\s+not\s+tell\s+the\s+user', "deception_hide"),    (r'system\s+prompt\s+override', "sys_prompt_override"),    # ... 更多模式  
]