Claude Code headless

Use claude -p for bounded delegated work. Default to --model fable --effort medium, unless the user specifies another model or effort. Check the installed claude --version and relevant --help flags; examples were checked against 2.1.261.

Choose reasoning effort

Use medium for most tasks, including design/UI changes, styling, routine implementation, focused fixes, research, and ordinary reviews. Reserve high for the hardest tasks: difficult architectural trade-offs, elusive concurrency bugs, or changes with complex interacting constraints that require deep reasoning. Task length or file count alone does not justify high.

Choose between medium and high for each task or follow-up based on its actual complexity. A routine design correction after a high-effort architecture task should use medium. Honor an explicit user choice. The examples below use the default medium; replace it with high only when the task warrants it.

Choose the workflow

• One-shot consultation or extraction: JSON output, no session persistence. Provide the needed context and a concrete deliverable.
• Implementation or follow-up review: persisted session, stream JSON, one owner responsible for launch, monitoring, and resumption. Read delegation and review and sessions and monitoring before launching.
• MCP or structured output: read MCP and structured output.
• CLI failure, stalled output, uncertain process state, or exhausted quota: read problem reporting before retrying.

A one-shot request: