Developer implementation guide for building hierarchical (folded) memory into an Agent. Three-layer architecture where recent turns stay detailed, older content compresses into episodes, and the oldest distills into durable semantic facts. Use when compact-memory-implementation is not retaining enough, or when agents need to recall decisions from many sessions ago.
Design and implement a layered, configurable permission/safety system for agent tools. Use this skill when building an agent that needs to control which tool calls are auto-allowed, which require user confirmation, and which are denied — especially when the system must be configurable across multiple scopes (project/user/enterprise) and extensible via hooks. Triggers on: "权限系统", "工具安全", "tool permission", "permission system", "tool safety", "allow/deny rules", "hook system", "构建安全机制".
Define agent tools using the fail-closed design pattern — unified name/schema/security/execution in one class, with three-layer execution (validate → permission → call). Use this skill whenever the user wants to define a new agent tool, add permission or validation logic to an existing tool, or asks about 'build a tool', '定义一个工具', 'create a tool for X', '工具定义'. Framework-agnostic: works with hermes-agent, LangChain, or any Python agent framework.
Build LangChain (Python) tools using Claude Code's fail-closed design pattern — unified name/schema/security/execution in one class, with automatic three-layer execution (validate → permission → call). Use this skill whenever the user wants to define a new LangChain tool, add permission or validation logic to an existing tool, set up the ClaudeStyleTool base class in a project, or asks about "build_tool", "Claude Code style tool", "工具定义", or "langchain tool with permissions". Also trigger when the user says "create a tool for X" or "定义一个工具" in a LangChain Python project context, even without mentioning Claude Code explicitly.
Developer implementation guide for adding compact memory to an Agent — covers fork agent pattern for compaction, trigger strategy, summary format design, and memory restoration in subsequent sessions. Use when a developer asks how to implement compact memory, context compression, or memory persistence in their agent built with Claude Agent SDK or Anthropic API.
Restructures a chaotic or overgrown MEMORY.md into a clean 2-layer architecture based on how Claude Code's autoDream system organizes memory — a lightweight pointer index (always loaded) and topic files (loaded on demand). Stale or superseded memories are deleted or corrected in place — not archived. Use this skill whenever the user says "clean up MEMORY.md", "reorganize my memory files", "MEMORY.md is getting too long", "fix my memory structure", or when you observe that MEMORY.md exceeds 200 lines, contains full paragraphs instead of pointers, or mixes index entries with topic content.
End-of-session memory distillation — extracts key decisions, eliminated approaches, new discoveries, and current blockers from the current conversation and writes them to MEMORY.md topic files. Based on Claude Code's autoDream background consolidation service. Activate when the user says "dream", "/dream", "save session memories", "distill this session", "what should I remember from this session", or when a long productive session is ending and the user wants to preserve what was learned.
Implement a production-ready LLM query loop / agent loop for AI applications. Use this skill whenever the user wants to add tool calling, ReAct-style reasoning-action-observation cycles, function calling loops, query engines, agent runtimes, tool_result feedback, max-turn exits, or Claude Code-like Agent Loop behavior to their own product or codebase.