WorkspaceBrowse, save, and manage patterns without leaving the shell

Claude Code Skills & Hooks Architecture

Full .claude/ directory with auto-discovered skills, slash commands, subagents, lifecycle hooks, and plugin support for advanced workflows.

heavyWeb AppFeatured
Q
Quzr27@quzr27

software engineer

1 followerquzr27
AI Tools
claude-code
File structure7 files
4
Pros
Hooks provide deterministic guarantees — safety rules always enforced
Skills auto-discover based on task context without manual invocation
Subagents isolate context-heavy tasks from the main session
Plugins bundle everything for team distribution
Cons
High setup complexity — requires understanding 5+ extension points
Skills are probabilistic — Claude may skip them during complex tasks
Custom subagents can gatekeep context from the main agent
Claude Code-only — none of this transfers to Cursor or Copilot
Avoid if
You are new to Claude Code — start with a simple CLAUDE.md first
Your project does not need automated quality gates or specialized workflows
You need portability across tools — skills, hooks, and subagents are Claude Code-only

Description

Claude Code's extensibility goes far beyond CLAUDE.md. Skills (.claude/skills/<name>/SKILL.md) are auto-discovered based on description matching — Claude loads them when relevant without manual invocation. Commands (.claude/commands/*.md) are explicit /slash triggers for repeatable workflows. Subagents (.claude/agents/*.md) get separate context windows with their own tool access and hooks. Hooks (in settings.json) are deterministic — they fire on PreToolUse, PostToolUse, Stop, and other lifecycle events with zero exceptions. The critical insight: skills are probabilistic (Claude decides when to use them), hooks are deterministic (always fire). Use hooks for safety rules, skills for conventions.

Who is this for

Power users and teams running Claude Code on production codebases who need automated quality gates, specialized workflows, and context isolation.

Why files are split this way

CLAUDE.md handles global context. Skills handle domain knowledge (e.g., how to generate PDFs, how to write tests). Commands handle repeatable workflows (e.g., /commit, /review-pr). Subagents handle tasks that need isolated context (e.g., research that would pollute the main window). Hooks handle non-negotiable rules (e.g., always run linter after edits). Each layer has a distinct role.

What changed from the previous setup

A large CLAUDE.md file trying to cover everything — conventions, workflows, safety rules, domain knowledge — in one place.