claude::code

curl -fsSL https://claude.ai/install.sh | bash
brew install --cask claude-code          # macOS
irm https://claude.ai/install.ps1 | iex  # Windows

# first session
cd your-project/
claude              # start interactive session
/init               # generate CLAUDE.md from codebase

# update
claude update

claude                    # interactive REPL
claude "explain this project"
claude -p "fix lint"       # print mode (one-shot)
cat err.log | claude -p "summarize"
claude -c                  # continue last conversation
claude -r my-session      # resume named session

claude auth login    # sign in
claude auth status   # check status
claude auth logout   # sign out

Persistent project instructions loaded automatically at start of every session.

/etc/claude-code/CLAUDE.md   # Managed policy (enterprise)
~/.claude/CLAUDE.md          # User (all projects)
./CLAUDE.md                  # Project root (committed)
./.claude/CLAUDE.md          # Alt project location
./src/module/CLAUDE.md       # Nested (on-demand)
./.claude/rules/*.md         # Path-scoped rules

~/.claude/projects/<hash>/memory/
  MEMORY.md      # Index (first 200 lines loaded)
  *.md           # Individual memory files

---
name: memory-name
description: "one-line summary"
type: user|feedback|project|reference
---

Content goes here.

✗ Code patterns / architecture (derive from code)

✗ Git history / who-changed-what (use git log)

✗ Debugging solutions (fix is in the code)

✗ Anything in CLAUDE.md already

✗ Ephemeral task details

Deterministic shell commands that run at lifecycle events. Defined in settings.json.

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Edit|Write",
      "hooks": [{
        "type": "command",
        "command": "my-script.sh"
      }]
    }]
  }
}

{
  "permissions": {
    "allow": [
      "Bash(git *)",
      "Bash(npm run *)",
      "Edit(/src/**/*.ts)",
      "WebFetch(domain:github.com)",
      "mcp__github__*"
    ],
    "deny": ["Bash(rm -rf *)"]
  }
}

Skills = custom slash commands. Claude auto-invokes via natural language match on description.

~/.claude/skills/my-skill/SKILL.md     # user-level
.claude/skills/my-skill/SKILL.md      # project-level

---
name: my-skill
description: "What it does"
user-invocable: true
allowed-tools: "Read,Grep,Bash(git *)"
effort: high
model: opus
context: fork          # run in subagent
paths: "src/**/*.ts"   # auto-load for files
shell: bash            # bash or powershell
---

Instructions for the skill.
Use $ARGUMENTS for user input.

## Live data
- Branch: !`git branch --show-current`
- Diff: !`git diff --stat`
- Status: !`git status -s`

Backtick commands in SKILL.md run at load time and inject output into the prompt.

Connect external tools & services. Claude can call MCP tools like native tools.

# In settings.json or .claude/mcp.json
{
  "mcpServers": {
    "github": {
      "type": "stdio",
      "command": "gh",
      "args": ["api-server"]
    },
    "remote": {
      "type": "http",
      "url": "https://mcp.example.com"
    }
  }
}

{
  "model": "opus",
  "effortLevel": "medium",
  "permissions": {
    "allow": ["Bash(git *)", "Edit"]
  },
  "hooks": { ... },
  "mcpServers": { ... },
  "env": { "DEBUG": "true" }
}

claude --model opus       # at startup
/model opus               # during session
Alt+P                    # keyboard shortcut

/effort low|medium|high|max
# max = Opus 4.6 only, deepest reasoning

Both IDEs support diff viewer, diagnostics sharing, and inline suggestions.

cd project && claude
  # 1. Describe feature / intent
  # 2. Let Claude explore (Plan mode)
  # 3. Review & approve implementation
/compact    # if context fills up
/clear      # between unrelated tasks
Esc Esc     # rewind if wrong direction

claude -w feature-auth   # isolated worktree 1
claude -w feature-ui     # isolated worktree 2
/batch "add tests to *"  # 5-30 parallel workers

claude -p "fix lint" --output-format json
cat err.log | claude -p "summarize"
claude -p "review" --max-turns 5

# use structure for complex requests
Goal: exploit SSTI safely in local challenge
Context: Flask/Jinja app in ./src
Constraints: no external tools
Files: @src/app.py
Expected output: curl PoC + explanation