AI 原生开放教材

外观

Claude How-To工程手册·英文原文 ↗原件 ↓仓库 ↗

Topic-Specific Recommendations ​

Use these specific recommendations when a topic is a gap. Paths are relative to this reference file (.claude/skills/self-assessment/references/).

Slash Commands (score 0):

  • Tutorial: 01-slash-commands/
  • Focus on: Built-in commands reference, creating your first SKILL.md, $ARGUMENTS syntax
  • Key exercise: Create a /optimize command and test it
  • Done when: You can create a custom skill with arguments and dynamic context

Slash Commands (score 1 — review):

  • Focus on: Dynamic context with !backtick`` syntax, @file references, disable-model-invocation vs user-invocable control
  • Done when: You can create a skill that injects live command output and controls its own invocation behavior

Memory (score 0):

  • Tutorial: 02-memory/
  • Focus on: CLAUDE.md creation, /init and /memory commands, # prefix for quick updates
  • Key exercise: Create a project CLAUDE.md with your coding standards
  • Done when: Claude remembers your preferences across sessions

Memory (score 1 — review):

  • Focus on: the 7 memory locations and how they are concatenated into context (loaded root-down, not overridden), .claude/rules/ directory with path-specific rules, @import syntax (max depth 4), Auto Memory MEMORY.md (Claude loads its first 200 lines or 25 KB, whichever comes first — a load cap, not a file-size limit)
  • Done when: You have modular rules for different directories and understand how every memory file is concatenated into context

Skills (score 0):

  • Tutorial: 03-skills/
  • Focus on: SKILL.md format, auto-invocation via description field, progressive disclosure (3 loading levels)
  • Key exercise: Install the code-review skill and verify it auto-triggers
  • Done when: A skill automatically activates based on conversation context

Skills (score 1 — review):

  • Focus on: context: fork with agent field for subagent execution, disable-model-invocation vs user-invocable, the skill-listing budget (1% of the context window, fallback 8,000 characters, 250 characters per entry), bundled resources (scripts/, references/, assets/)
  • Done when: You can create a skill that runs in a subagent with forked context

Hooks (score 0):

  • Tutorial: 06-hooks/
  • Focus on: Configuration structure (matcher + hooks array), PreToolUse/PostToolUse events, exit codes (0=success, 2=block), JSON input/output format
  • Key exercise: Create a PreToolUse hook that validates Bash commands
  • Done when: A hook blocks dangerous commands before execution

Hooks (score 1 — review):

  • Focus on: All 33 hook events (including PostToolUseFailure, StopFailure, TaskCreated, CwdChanged, FileChanged, PostCompact, Elicitation, ElicitationResult, Setup, UserPromptExpansion, MessageDisplay, PreModelSwitch, PostModelSwitch — the last two added in v2.1.251), 5 hook types (command, http, mcp_tool, prompt, agent — agent hooks are experimental and may change), component-scoped hooks in SKILL.md frontmatter, HTTP hooks with allowedEnvVars, CLAUDE_ENV_FILE for SessionStart/CwdChanged/FileChanged
  • Done when: You can create a prompt-based Stop hook and a component-scoped hook in a skill

MCP (score 0):

  • Tutorial: 05-mcp/
  • Focus on: claude mcp add command, transport types (http recommended, stdio, ws for push-style servers, and the deprecated sse — note --transport does not accept ws, so add WebSocket servers with claude mcp add-json), GitHub MCP setup, environment variable expansion
  • Key exercise: Add GitHub MCP server and query PRs
  • Done when: You can query live data from an external service via MCP

MCP (score 1 — review):

  • Focus on: Project-scope .mcp.json (requires team approval), OAuth 2.0 auth, MCP resources with @server:resource mentions, Tool Search (ENABLE_TOOL_SEARCH), claude mcp serve, output limits (10,000 tokens warning; 25,000 tokens default max via MAX_MCP_OUTPUT_TOKENS; 50,000 characters disk-persistence threshold)
  • Done when: You have a project .mcp.json and understand Tool Search auto mode

Subagents (score 0):

  • Tutorial: 04-subagents/
  • Focus on: Agent file format (.claude/agents/*.md), built-in agents (Explore, Plan, general-purpose, claude, statusline-setup, claude-code-guide), tools/model/permissionMode config, spawn limits (depth default 3 since v2.1.219, concurrency default 20, and the 200-per-session spawn cap removed in v2.1.224)
  • Key exercise: Create a code-reviewer subagent and test delegation
  • Done when: Claude delegates code review to your custom agent

Subagents (score 1 — review):

  • Focus on: Worktree isolation (isolation: worktree), persistent agent memory (memory field with scopes), background agents (Ctrl+B/Ctrl+F), agent allowlists with Agent(agent_type) (Task(...) remains a back-compat alias), agent teams (--teammate-mode)
  • Done when: You have a subagent with persistent memory running in worktree isolation

Checkpoints (score 0):

  • Tutorial: 08-checkpoints/
  • Focus on: Esc+Esc and /rewind access, 6 rewind options (restore code and conversation, restore conversation, restore code, summarize from here, summarize up to here, never mind), limitations — bash filesystem ops, subagent edits (except a foreground context: fork skill), edits made outside Claude Code, and symlinked/hardlinked paths are all untracked
  • Key exercise: Make experimental changes, then rewind to restore
  • Done when: You can confidently experiment knowing you can rewind

Advanced Features (score 0):

  • Tutorial: 09-advanced-features/
  • Focus on: Planning mode (/plan or Shift+Tab), permission modes (6 types: manual — renamed from default in v2.1.200 — acceptEdits, plan, auto, dontAsk, bypassPermissions), extended thinking (Alt+T toggle)
  • Key exercise: Use planning mode to design a feature, then implement it
  • Done when: You can switch between planning and implementation modes fluently

Advanced Features (score 1 — review):

  • Focus on: Remote control (claude --remote-control, alias --rc), web sessions (claude --cloud; --remote is a deprecated alias), desktop handoff (/desktop), worktrees (claude -w), task lists (Ctrl+T), managed settings for enterprise
  • Done when: You can hand off sessions between CLI, web, and desktop

Plugins (score 0):

  • Tutorial: 07-plugins/
  • Focus on: Plugin structure (.claude-plugin/plugin.json), what plugins bundle (skills, agents, MCP, hooks, settings — plus the legacy commands/ directory, which still works but skills/ is preferred for new plugins), installation from marketplace
  • Key exercise: Install a plugin and explore its components
  • Done when: You understand when to use a plugin vs standalone components

Plugins (score 1 — review):

  • Focus on: Creating plugin.json manifest, plugin hooks (hooks/hooks.json), LSP configuration (.lsp.json), ${CLAUDE_PLUGIN_ROOT\} variable, --plugin-dir for testing, marketplace publishing
  • Done when: You can create and test a plugin for your team

CLI (score 0):

  • Tutorial: 10-cli/
  • Focus on: Interactive vs print mode, claude -p with piping, --output-format json, session management (-c/-r)
  • Key exercise: Pipe a file to claude -p and get JSON output
  • Done when: You can use Claude non-interactively in a script

CLI (score 1 — review):

  • Focus on: --agents flag with JSON config, --json-schema for structured output, --fallback-model, --from-pr, --strict-mcp-config, batch processing with for loops, claude mcp serve
  • Done when: You have a CI/CD script that uses Claude with structured JSON output

Last Updated: September 2, 2026 Claude Code Version: 2.1.257 Sources: