Claude CLI Skill

Run Claude Code CLI to leverage Anthropic's models via OAuth authentication (no API tokens required).

When to Use

• Complex coding tasks that benefit from extended context
• Multi-file code generation or refactoring
• Research and analysis requiring deep reasoning
• Tasks that would benefit from Claude's thinking mode
• Delegating substantial work to a sub-process

Quick Start

One-shot query (non-interactive)

echo "Your prompt here" | claude -p

With specific model

echo "Complex task requiring Opus" | claude -p --model opus
echo "Standard task with Sonnet" | claude -p --model sonnet

Note: Piping the prompt via stdin is more reliable than passing it as an argument.

With thinking/extended reasoning

claude -p --model opus "Analyze this complex problem"

JSON output (for parsing)

claude -p --output-format json "Your prompt"

Stream JSON (real-time)

claude -p --output-format stream-json "Your prompt"

Key Options

Working with Files

Claude CLI can read and write files in its working directory:

# Analyze a file
claude -p "Review this code for bugs" < myfile.py

# Work in a specific directory
cd /path/to/project && claude -p "Summarize this codebase"

# Give access to additional directories
claude -p --add-dir /other/path "Work with files in both locations"

For Sub-Agent Delegation

When spawning Claude CLI for background work:

# Run with timeout protection
timeout 300 claude -p --model sonnet "Complete this task..." 2>&1

# Capture structured output
claude -p --output-format json "Generate a report" > result.json

Script: Run Claude Task

Use the bundled script for reliable sub-agent execution:

node {baseDir}/scripts/run-claude-task.cjs "Your task prompt" [--model sonnet|opus] [--timeout 300]

The script handles:

• Timeout protection
• Error capture and formatting
• Clean output for parsing
• Exit code propagation

Authentication

Claude CLI uses OAuth by default (linked to your Anthropic account). No API key configuration needed.

To check auth status:

claude --version

If not authenticated, run interactively once:

claude

Follow the OAuth prompts to link your account.

Model Selection

Tips

1. Always use -p for non-interactive/scripted use
2. Set timeouts for long-running tasks to prevent hangs
3. Use --output-format json when you need to parse results
4. Pipe input for file analysis: cat file.py | claude -p "Review this"
5. Working directory matters — Claude sees files relative to cwd

Limitations

• OAuth requires initial interactive setup
• Rate limits apply based on your Anthropic account tier
• Some operations may timeout on very complex tasks
• Interactive features (like file editing confirmations) don't work with -p