# Claude Code 智能体编码工具详解：25 项功能与策略 - 来源：MarkTechPost（RSS） - 作者：Michal Sutter - 发布时间：2026-06-15 09:04 - AIHOT 分数：61 - AIHOT 链接：https://aihot.virxact.com/items/cmqeiikyy02k9slunbwtuqblk - 原文链接：https://www.marktechpost.com/2026/06/14/claude-code-guide-2026-25-features-with-examples-demo ## AI 摘要 Claude Code 是 Anthropic 的智能体编码工具，运行于终端、桌面应用和 IDE，基于智能体循环工作。文章将 25 项功能与策略分为官方功能、社区技术和第三方工具三类，并逐一标注。官方功能包括 CLAUDE.md 记忆文件、技能、子智能体、斜杠命令（/init、/compact、/review 等）、钩子、MCP 服务器、插件、检查点、计划模式、权限模式、自动模式（使用 Sonnet 4.6 分类器）、上下文压缩、后台任务、Agent SDK、无头 CLI、GitHub Action 集成、输出样式、远程控制与移动推送、离线摘要、沙盒。社区技术涵盖结构化上下文文件夹、动态工作流、模块化技能管道、弹性技术。第三方工具如 Mem Search 可扩展外部记忆层。 ## 正文 Claude Code started as a terminal coding assistant. It now runs as a layered agentic system. Underneath, Claude Code separates memory, hooks, skills, subagents, plugins, and MCP into distinct layers. Each layer changes what the model can see or do. This article covers 25 features and strategies for scaling Claude Code. It is written for AI engineers, software engineers, and data scientists. Every code example follows a documented format and runs as written. Each item is labeled by status, so you know what ships with Claude Code and what does not. What is Claude Code Claude Code is Anthropic’s agentic coding tool. It works in the terminal, the desktop app, and your IDE. It can read files, run commands, edit code, and call external tools. Under the hood, it runs an agentic loop. That loop chooses tools, accumulates context, and manages long sessions through compaction. Safety boundaries come from permission modes, checkpoints, sandboxing, and managed settings. The same loop is exposed programmatically through the Agent SDK. Developers extend the tool with a small set of primitives. Those primitives are CLAUDE.md, skills, subagents, slash commands, hooks, and MCP servers. Plugins bundle these primitives into one installable unit. The 25 Features and Strategies Each feature/strategy is labeled. ‘Official’ means documented Anthropic functionality. ‘Community technique’ means a workflow pattern, not a shipped feature. ‘Third-party tool’ means software built outside Anthropic. CLAUDE.md memory file (Official). This file is the agent’s constitution for your repository. Claude reads it every session to anchor conventions and commands. Skills (Official). A skill is a SKILL.md file with frontmatter under .claude/skills//. It supports /name invocation and autonomous invocation by Claude. SKILL.md .claude/skills// /name Subagents (Official). Subagents are specialized instances with their own context windows. Verbose work stays isolated, so your main conversation stays focused. Slash commands (Official). These are typed shortcuts starting with /. Built-ins include /init, /compact, /context, /review, and /security-review. / /init /compact /context /review /security-review Hooks (Official). Hooks are deterministic scripts that fire at defined lifecycle points. PreToolUse is the primary security checkpoint before any tool runs. PreToolUse MCP servers (Official). Model Context Protocol connects Claude Code to GitHub, databases, and browsers. The server handles integration; Claude reasons about what to do. Plugins (Official). A plugin is a versioned bundle of skills, subagents, commands, hooks, and MCP definitions. One /plugin command installs the whole set. /plugin Checkpoints (Official). Claude Code snapshots state automatically before changes. Press Escape twice to rewind when something breaks. Plan mode (Official). Plan mode explores and proposes without executing. It is ideal for scoping work before committing edits. Permission modes (Official). Default mode asks before each file write and shell command. Other modes trade oversight for speed. Auto Mode (Official, research preview). A separate Sonnet 4.6 classifier reviews each action first. Safe actions proceed; risky ones get blocked or escalated. Context compaction (Official). /compact condenses long sessions to preserve usable context. /context reports current context usage. /compact /context Background tasks (Official). Long shell commands run with the run_in_background flag on the Bash tool. Claude polls output without blocking the conversation. run_in_background Agent SDK (Official). The SDK exposes the same loop programmatically through query(). You can send slash commands like /code-review and process results. query() /code-review Headless CLI (Official). claude -p "query" runs a one-shot process and exits. Piped input like cat logs.txt | claude -p also works. claude -p "query" cat logs.txt | claude -p GitHub Action and scheduled jobs (Official). Claude Code runs as a one-shot process without a TTY. This enables CI integration, scheduled jobs, and pre-commit hooks. Output styles and statusLine (Official). Output styles change response formatting. A custom statusLine renderer surfaces session state in the terminal. Remote Control and mobile push (Official). You can drive Claude Code from mobile or web surfaces. Claude can send push notifications when tasks finish. Away summary (Official). This session-level feature surfaces context when you return to a paused session. It is enabled by default and can be opted out. Sandboxing (Official). The sandboxed Bash tool enforces OS-level filesystem and network isolation. Commands run without prompts inside boundaries you define. Structured context folders (Community technique). Organize task-specific folders for brand guidelines, client data, or legal terminology. The right context loads for each task, improving output relevance. Dynamic workflows (Community technique). Break complex tasks into smaller steps using sub-agents. Common patterns include ‘classify and act’ and ‘fan out and synthesize.’ Modular skill pipelines (Community technique). Chain reusable skills into end-to-end workflows. A support pipeline can combine categorization, response generation, and escalation skills. External memory layers (Third-party tool). Tools such as Mem Search or Hermes extend recall across long projects. They sit outside Claude Code’s built-in memory. Resilience techniques (Community technique). Practitioners reset and retry tasks when output quality degrades. This avoids context pollution and keeps results consistent. Try the Interactive Demo /help /init /context /review /security-review /mcp /agents /compact How the Extensibility Primitives Compare Devs/AI Professionals often confuse skills, subagents, slash commands, and hooks. The table below separates them by where they live and how they run. PrimitiveWhere it livesHow it runsIsolated context?Best forSlash command.claude/commands/ (legacy)Typed /nameNoInserting a prompt templateSkill.claude/skills//SKILL.md/name or autonomousOptional (via subagent)Domain logic with shipped filesSubagent.claude/agents/Auto-delegate or @agent-nameYes, own context windowIsolated, parallel tasksHookSettings, skill, or subagent frontmatterEvent-driven at lifecycle pointsRuns deterministic codeEnforcing rules without hallucinationMCP server.mcp.json or claude mcp addTool calls to a serverExternal processGitHub, databases, browsersPluginInstalled via /pluginBundles all of the aboveInherits component scopeSharing setups across teams .claude/commands/ /name .claude/skills//SKILL.md /name .claude/agents/ @agent-name .mcp.json claude mcp add /plugin The rule of thumb is simple. Use a slash command for a prompt template. Use a skill when there is real domain logic or helper files. Use a subagent for isolated, parallel work. Use a hook to enforce a rule with code. Use Cases With Examples Codebase onboarding: Join a new team and run an Explore subagent. It is read-only, so it maps the repo without editing files. Pair it with a CLAUDE.md that lists build, lint, and test commands. Automated code review: Run /review for general feedback or /security-review for vulnerabilities. On Team and Enterprise plans, multi-agent review can split the work across subagents. /review /security-review Overnight refactors: Enable Auto Mode for a clearly scoped task in an isolated environment. Combine it with background tasks and checkpoints. If output drifts, rewind with Escape twice and retry. Customer feedback classification: Build a dynamic ‘classify and act’ workflow. Claude reads feedback, categorizes responses, and generates insights in one pass. This suits high-volume, repetitive operations. Continuous integration. Use the headless CLI inside a GitHub Action. Run claude -p on each pull request to lint, test, or summarize diffs without a terminal. Scheduled jobs can run the same command nightly. Combine this with hooks to enforce team policy automatically. claude -p Coding Examples These snippets are illustrative and faithful to documented formats. Start with a minimal CLAUDE.md. # Project: my-tool ## Build npm run build ## Test npm test ## Conventions - TypeScript strict mode - No default exports - Commit format: feat/fix/chore(scope): description A skill is a folder with a SKILL.md and frontmatter. SKILL.md --- name: code-review description: Review changed files against our team standards. --- Review staged changes. Flag risks. Suggest concrete fixes. A subagent restricts its tools to stay read-only and safe. A subagent restricts its tools to stay read-only and safe. --- name: explorer description: Read-only codebase exploration. tools: Read, Grep, Glob --- Map the repository structure and summarize entry points. A PreToolUse hook blocks dangerous Bash before it runs. PreToolUse { "hooks": { "PreToolUse": [ { "matcher": "Bash", "hooks": [ { "type": "command", "command": "scripts/guard.sh" } ] } ] } } Add a maintained MCP server, then run a headless query for CI. claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/projects claude -p "List the largest files under src and explain why" --output-format json Key Takeaways Claude Code is a layered agentic system, not a single chat prompt. Six primitives drive extensibility: CLAUDE.md, skills, subagents, slash commands, hooks, and MCP. Auto Mode is a research-preview permission mode gated by a Sonnet 4.6 classifier. Not every ‘Claude Code’ tip is official; some rely on third-party tools. Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights. Michal SutterPerplexity Moves Deep Research Into Computer, Routing Research Subtasks Across 20+ Frontier Models For Reports, Decks, And Dashboards Michal SutterxAI Ships Grok Build Plugin Marketplace With MongoDB, Vercel, Sentry, Chrome DevTools, Cloudflare, and Superpowers Plugins at Launch Michal SutterNous Research Ships Hermes Agent Profile Builder: Identity, Model, Skills, and MCP Servers in One Dashboard Flow Michal SutterTop AI Coding Agents and Development Platforms in 2026: Atoms, Devin, Windsurf, Cursor, Warp, and More Compared Michal SutterGoogle Research Adds Agentic RAG to Gemini Enterprise Agent Platform with a Sufficient Context Agent for multi-hop queries Michal SutterBest 21 Low-Code and No-Code AI Tools in 2026 Michal SutterMoonshot AI Releases Kimi Code CLI: A Terminal AI Coding Agent Built in TypeScript for Next-Gen Agents Michal SutterPerplexity AI Introduces Hybrid Local-Server Inference Orchestrator for Personal Computer: Automatic On-Device and Cloud Task Routing Michal SutterNous Research Releases Hermes Desktop: A Native Cross-Platform Front End for Hermes Agent v0.15.2 with Streaming Tool Output Michal SutterAlibaba’s Qwen Team Launches Qwen3.7-Plus, Adding Vision, Deep Reasoning, Tool Invocation, and Autonomous Iteration on the Bailian Platform Michal SutterMeet Memory OS: A 6-Layer Open-Source Memory Stack Built on Top of Hermes Agent Michal SutterTrajectory Releases a Concurrent Multi-LoRA Training Stack for Continual Learning, Reporting a 2.81× Experiment-Throughput Gain Michal SutterGenesis AI Releases Nyx, Quadrants, and Genesis World 1.0 Physics Platform for Scalable Robotics Foundation Model Evaluation Michal SutterAnthropic Ships Claude Opus 4.8 Alongside Dynamic Workflows and Cheaper Fast Mode, With Workflows Capped at 1,000 Subagents Michal SutterMeet EAGLE 3.1: The Speculative Decoding Algorithm That Fixes Attention Drift in LLM Inference Michal SutterMeet OmniVoice Studio: A Local, Open-Source Alternative to ElevenLabs Michal SutterStepFun Releases StepAudio 2.5 Realtime: An End-to-End Voice Model with Roleplay-Specific RLHF and Paralinguistic Comprehension Michal SutterTencent Open-Sources TencentDB Agent Memory: A 4-Tier Local Memory Pipeline for AI Agents Michal SutterCohere Releases Command A+: A 218B Sparse MoE Model for Agentic Workflows That Runs on as Few as Two H100 GPUs Michal SutterWhat is a Forward Deployed Engineer: The AI Role OpenAI, Anthropic, and Google Are Hiring in 2026 Michal SutterGoogle Introduces Gemini 3.5 Flash at I/O 2026: A Faster and Cheaper Model for AI Agents and Coding Michal SutterUpstash for Redis vs Supabase vs Neon: Which One Fits Vibe Coding Workflows in 2026? Michal SutterGoogle Launches Antigravity 2.0 at I/O 2026: A Standalone Agent-First Platform with CLI, SDK, Managed Execution, and Enterprise Support Michal SutterVercel Labs Introduces Zero, a Systems Programming Language Designed So AI Agents Can Read, Repair, and Ship Native Programs Michal SutterEnterprise AI Governance in 2026: Why the Tools Employees Use Are Ahead of the Policies That Cover Them Michal SutterGoogle DeepMind Introduces an AI-Enabled Mouse Pointer Powered by Gemini That Captures Visual and Semantic Context Around the Cursor Michal SutterOpenAI Introduces Daybreak: A Cybersecurity Initiative That Puts Codex Security at the Center of Vulnerability Detection and Patch Validation Michal SutterBest Vector Databases in 2026: Pricing, Scale Limits, and Architecture Tradeoffs Across Nine Leading Systems Michal SutterOpenClaw vs Hermes Agent: Why Nous Research’s Self-Improving Agent Now Leads OpenRouter’s Global Rankings Michal SutterNVIDIA AI Just Released cuda-oxide: An Experimental Rust-to-CUDA Compiler Backend that Compiles SIMT GPU Kernels Directly to PTX Michal SutterOpenAI Introduces MRC (Multipath Reliable Connection): A New Open Networking Protocol for Large-Scale AI Supercomputer Training Clusters Michal SutterGoogle Adds Event-Driven Webhooks to the Gemini API, Eliminating the Need for Polling in Long-Running AI Jobs Michal SutterMicrosoft Research’s World-R1 Uses Flow-GRPO and 3D-Aware Rewards to Inject Geometric Consistency Into Wan 2.1 Without Architectural Changes Michal SutterCursor Introduces a TypeScript SDK for Building Programmatic Coding Agents With Sandboxed Cloud VMs, Subagents, Hooks, and Token-Based Pricing Michal SutterTop 10 KV Cache Compression Techniques for LLM Inference: Reducing Memory Overhead Across Eviction, Quantization, and Low-Rank Methods Michal Suttersmol-audio: A Colab-Friendly Notebook Collection for Fine-Tuning Whisper, Parakeet, Voxtral, Granite Speech, and Audio Flamingo 3 Michal SutterxAI Launches grok-voice-think-fast-1.0: Topping τ-voice Bench at 67.3%, Outperforming Gemini, GPT Realtime, and More Michal SutterGoogle DeepMind Introduces Vision Banana: An Instruction-Tuned Image Generator That Beats SAM 3 on Segmentation and Depth Anything V3 on Metric Depth Estimation Michal SutterOpenAI Releases GPT-5.5, a Fully Retrained Agentic Model That Scores 82.7% on Terminal-Bench 2.0 and 84.9% on GDPval Michal SutterNext Leap to Harness Engineering: JiuwenClaw Pioneers ‘Coordination Engineering’ Michal SutterOpenAI Scales Trusted Access for Cyber Defense With GPT-5.4-Cyber: a Fine-Tuned Model Built for Verified Security Defenders Michal SutterxAI Launches Standalone Grok Speech-to-Text and Text-to-Speech APIs, Targeting Enterprise Voice Developers Michal SutterA Coding Tutorial for Running PrismML Bonsai 1-Bit LLM on CUDA with GGUF, Benchmarking, Chat, JSON, and RAG Michal SutterTop 19 AI Red Teaming Tools (2026): Secure Your ML Models Michal SutterA Coding Guide to Build a Production-Grade Background Task Processing System Using Huey with SQLite, Scheduling, Retries, Pipelines, and Concurrency Control Michal SutterGoogle AI Launches Gemini 3.1 Flash TTS: A New Benchmark in Expressive and Controllable AI Voice Michal SutterA Coding Implementation of Crawl4AI for Web Crawling, Markdown Generation, JavaScript Execution, and LLM-Based Structured Extraction Michal SutterGoogle AI Research Proposes Vantage: An LLM-Based Protocol for Measuring Collaboration, Creativity, and Critical Thinking Michal SutterMeta AI and KAUST Researchers Propose Neural Computers That Fold Computation, Memory, and I/O Into One Learned Model Michal SutterA Coding Implementation of MolmoAct for Depth-Aware Spatial Reasoning, Visual Trajectory Tracing, and Robotic Action Prediction Michal SutterAlibaba’s Tongyi Lab Releases VimRAG: a Multimodal RAG Framework that Uses a Memory Graph to Navigate Massive Visual Contexts Michal SutterA Coding Guide to Markerless 3D Human Kinematics with Pose2Sim, RTMPose, and OpenSim Michal SutterA Comprehensive Implementation Guide to ModelScope for Model Search, Inference, Fine-Tuning, Evaluation, and Export Michal SutterHow to Combine Google Search, Google Maps, and Custom Functions in a Single Gemini API Call With Context Circulation, Parallel Tool IDs, and Multi-Step Agentic Chains Michal SutterHow to Deploy Open WebUI with Secure OpenAI API Integration, Public Tunneling, and Browser-Based Chat Access Michal SutterNetflix AI Team Just Open-Sourced VOID: an AI Model That Erases Objects From Videos — Physics and All Michal SutterGoogle DeepMind’s Research Lets an LLM Rewrite Its Own Game Theory Algorithms — And It Outperformed the Experts Michal SutterHugging Face Releases TRL v1.0: A Unified Post-Training Stack for SFT, Reward Modeling, DPO, and GRPO Workflows Michal SutterGoogle AI Releases Veo 3.1 Lite: Giving Developers Low Cost High Speed Video Generation via The Gemini API Michal SutterAgent-Infra Releases AIO Sandbox: An All-in-One Runtime for AI Agents with Browser, Shell, Shared Filesystem, and MCP Michal SutterGoogle-Agent vs Googlebot: Google Defines the Technical Boundary Between User Triggered AI Access and Search Crawling Systems Today Michal SutterA Coding Guide to Exploring nanobot’s Full Agent Pipeline, from Wiring Up Tools and Memory to Skills, Subagents, and Cron Scheduling Michal SutterAn Implementation of IWE’s Context Bridge as an AI-Powered Knowledge Graph with Agentic RAG, OpenAI Function Calling, and Graph Traversal Michal SutterMeta Releases TRIBE v2: A Brain Encoding Model That Predicts fMRI Responses Across Video, Audio, and Text Stimuli Michal SutterTencent AI Open Sources Covo-Audio: A 7B Speech Language Model and Inference Pipeline for Real-Time Audio Conversations and Reasoning Michal SutterA Coding Implementation to Design Self-Evolving Skill Engine with OpenSpace for Skill Learning, Token Efficiency, and Collective Intelligence Michal SutterLuma Labs Launches Uni-1: The Autoregressive Transformer Model that Reasons through Intentions Before Generating Images Michal SutterMeet GitAgent: The Docker for AI Agents that is Finally Solving the Fragmentation between LangChain, AutoGen, and Claude Code Michal SutterA Coding Implementation for Building and Analyzing Crystal Structures Using Pymatgen for Symmetry Analysis, Phase Diagrams, Surface Generation, and Materials Project Integration Michal SutterA Coding Implementation Showcasing ClawTeam’s Multi-Agent Swarm Orchestration with OpenAI Function Calling Michal SutterA Coding Guide to Implement Advanced Differential Equation Solvers, Stochastic Simulations, and Neural Ordinary Differential Equations Using Diffrax and JAX Michal SutterBaidu Qianfan Team Releases Qianfan-OCR: A 4B-Parameter Unified Document Intelligence Model Michal SutterGoogle AI Releases WAXAL: A Multilingual African Speech Dataset for Training Automatic Speech Recognition and Text-to-Speech Models Michal SutterLangChain Releases Deep Agents: A Structured Runtime for Planning, Memory, and Context Isolation in Multi-Step AI Agents Michal SutterGoogle DeepMind Introduces Aletheia: The AI Agent Moving from Math Competitions to Fully Autonomous Professional Research Discoveries Michal SutterGoogle AI Introduces ‘Groundsource’: A New Methodology that Uses Gemini Model to Transform Unstructured Global News into Actionable, Historical Data Michal SutterHow to Build a Self-Designing Meta-Agent That Automatically Constructs, Instantiates, and Refines Task-Specific AI Agents Michal SutterA Coding Guide to Build a Complete Single Cell RNA Sequencing Analysis Pipeline Using Scanpy for Clustering Visualization and Cell Type Annotation Michal SutterHow to Build Progress Monitoring Using Advanced tqdm for Async, Parallel, Pandas, Logging, and High-Performance Workflows Michal SutterGoogle Launches TensorFlow 2.21 And LiteRT: Faster GPU Performance, New NPU Acceleration, And Seamless PyTorch Edge Deployment Upgrades Michal SutterOpenAI Introduces Codex Security in Research Preview for Context-Aware Vulnerability Detection, Validation, and Patch Generation Across Codebases Michal SutterA Coding Guide to Build a Scalable End-to-End Machine Learning Data Pipeline Using Daft for High-Performance Structured and Image Data Processing Michal SutterHow to Build an EverMem-Style Persistent AI Agent OS with Hierarchical Memory, FAISS Vector Retrieval, SQLite Storage, and Automated Memory Consolidation Michal SutterMeet NullClaw: The 678 KB Zig AI Agent Framework Running on 1 MB RAM and Booting in Two Milliseconds Michal SutterHow to Build an Explainable AI Analysis Pipeline Using SHAP-IQ to Understand Feature Importance, Interaction Effects, and Model Decision Breakdown Michal SutterA Complete End-to-End Coding Guide to MLflow Experiment Tracking, Hyperparameter Optimization, Model Evaluation, and Live Model Deployment