---

name: claude-cli description: Invoke Claude Code CLI directly with prompts using non-interactive mode

Claude CLI Skill

This skill enables you to invoke the Claude Code CLI directly with prompts in non-interactive mode, bypassing the workspace trust dialog.

Overview

The Claude Code CLI can be invoked in non-interactive mode using the -p or --print flag, which prints the response and exits. This is particularly useful for:

• Scripting and automation
• Piping output to other commands
• Quick one-off queries
• Integration with other tools

Important Security Note: The -p flag automatically bypasses the workspace trust dialog. Only use this in directories you trust.

OpenClaw Execution

When calling this skill from an OpenClaw agent, ALWAYS use pty:true to ensure the terminal output is handled correctly.

# Example OpenClaw tool call
bash pty:true command:"claude -p 'Your prompt'"

Automated Permission Approval

When using this skill for complex automation (e.g., orchestrating other Claude agents), you may encounter permission prompts that block the non-interactive flow. To pre-approve specific tools and commands in a project directory, create a .claude/settings.json file.

Common Language Permissions

Add these to the allow array based on the project's tech stack:

• General: Read, Write, Edit, Bash(*)
• Rust: Bash(cargo*), Bash(rustc*)
• Node/JS: Bash(npm*), Bash(pnpm*), Bash(bun*), Bash(node*)
• Python: Bash(python*), Bash(pip*), Bash(pytest*), Bash(poetry*)
• Go: Bash(go*)

Example for a multi-stack project:

mkdir -p .claude
cat > .claude/settings.json << 'EOF'
{
  "permissions": {
    "allow": [
      "Read", "Write", "Edit",
      "Bash(cargo*)", "Bash(rustc*)",
      "Bash(npm*)", "Bash(pnpm*)", "Bash(bun*)", "Bash(node*)",
      "Bash(python*)", "Bash(pip*)", "Bash(go*)"
    ]
  }
}
EOF

Troubleshooting & Unblocking Sub-agents

If an orchestrator agent spawns a background sub-agent that seems stuck:

1. Check the Logs: Use process action:log sessionId:XXX to look for errors like "blocked on permissions", "Not logged in", or "Please run /login".
2. Inject Permissions: If blocked on tools, use mkdir -p .claude and write the settings.json file as shown above.
3. Verify Auth: If "Not logged in", the human must run claude /login in a terminal once.
4. Restart: After fixing permissions/auth, the orchestrator should restart the sub-agent task.

Usage

Basic Syntax

claude -p "<prompt>"

Key Flags

• -p, --print: Print response and exit (non-interactive mode). This automatically skips the workspace trust dialog.
• --model <model>: Specify a model (e.g., 'sonnet', 'opus', or full model name)
• --output-format <format>: Output format - "text" (default), "json", or "stream-json"
• --max-budget-usd <amount>: Set maximum spending limit for API calls
• --session-id <uuid>: Use a specific session ID
• --system-prompt <prompt>: Override the system prompt
• --tools <tools>: Specify available tools (e.g., "Bash,Edit,Read" or "" to disable all)

Examples

Simple Query

claude -p "What is the capital of France?"

Code Analysis

claude -p "Analyze the code in main.py and suggest improvements"

With Model Selection

claude -p --model opus "Explain quantum computing in simple terms"

JSON Output

claude -p --output-format json "List the top 3 programming languages"

With Budget Limit

claude -p --max-budget-usd 0.50 "Generate a Python script to parse CSV files"

Disable All Tools (API-only mode)

claude -p --tools "" "Explain the concept of recursion"

With Custom System Prompt

claude -p --system-prompt "You are a code reviewer" "Review this function for security issues"

Integration Patterns

Piping Output

claude -p "Generate a bash script to backup files" > backup.sh
chmod +x backup.sh

Chaining Commands

result=$(claude -p "What is 2+2?")
echo "Claude says: $result"

Processing Files

for file in *.py; do
  claude -p "Summarize this Python file: $file"
done

When to Use This Skill

Use this skill when you need to:

1. Automate Claude interactions - Scripts, CI/CD pipelines, or automated workflows
2. Get quick answers - One-off queries without starting an interactive session
3. Process output programmatically - Pipe to files, parse JSON, or chain with other commands
4. Bypass trust prompts - Work in trusted directories without manual confirmation
5. Integrate with other tools - Combine Claude with existing command-line workflows

Security Considerations

⚠️ Warning: The -p flag bypasses the workspace trust dialog. This means:

• Claude will have access to tools without prompting for permission
• Only use this in directories you trust
• Be cautious when running in directories with sensitive data
• Consider using --tools "" to disable all tools if you only need text generation

Advanced Usage

Streaming JSON Output

claude -p --output-format stream-json "Explain machine learning" | jq -r '.content'

Fallback Model (when primary is overloaded)

claude -p --fallback-model haiku "Quick question about Python syntax"

With MCP Servers

claude -p --mcp-config ./mcp-config.json "Query the database for user stats"

Structured Output with JSON Schema

claude -p --json-schema '{"type":"object","properties":{"name":{"type":"string"},"age":{"type":"number"}},"required":["name","age"]}' "Extract person info: John is 30 years old"

Tips and Best Practices

1. Quote your prompts: Always use quotes around prompts to handle spaces and special characters
2. Use specific models: Specify --model for consistent behavior (e.g., --model sonnet for faster responses)
3. Set budget limits: Use --max-budget-usd to prevent unexpected API costs
4. Disable tools when not needed: Use --tools "" for pure text generation to improve speed and reduce costs
5. Capture output: Redirect output to files or variables for further processing
6. Check exit codes: The command returns 0 on success, non-zero on errors

Troubleshooting

Command not found

Ensure Claude CLI is installed and in your PATH:

which claude

Permission errors

Make sure you're in a directory you trust, or the -p flag won't help with file access permissions.

API rate limits

Use --max-budget-usd to control spending, or wait if you hit rate limits.

Related Commands

• claude --help: View all available options
• claude --version: Check installed version
• claude doctor: Check CLI health and configuration
• claude setup-token: Configure authentication