writing-claude-commands

Keep Claude commands minimal and focused: 5-15 lines with 1-3 steps for direct tasks, or reference a dedicated skill for complex workflows. This ensures commands remain fast entry points rather than overly engineered prompts that slow down execution.

Commands are simple, thin entry points with 1-3 steps and direct instructions or skill references. Skills handle multi-phase workflows with 4+ steps. Use the Simplicity Test and step count to distinguish: if your prompt exceeds 3 steps or needs personas and complex output templates, build a skill instead.

No. Claude commands enforce constraints that disallow personas, multi-phase workflows, and overlong outputs. These elements trigger a shift to skill-based design, which handles complexity through structured, reusable workflows separate from quick commands.

Claude commands require frontmatter metadata and follow pattern templates for direct instructions, skill wrappers, or tool permissions. Structure depends on complexity: simple commands use 5-15 lines with clear steps, while tool permissions and delegation patterns vary by use case.

Use step count and complexity as your guide: 1-3 steps with minimal reasoning = command. 4+ steps, multiple phases, or intricate workflows = skill. Commands act as entry points; skills handle the heavy lifting. If your prompt needs checklists, personas, or extended logic, delegate to a skill.

Three core patterns: direct-instruction commands for simple tasks, skill-wrapper commands that invoke dedicated workflows, and tool-permission commands that manage API access. Each pattern uses 5-15 lines and enforces the 1-3 step limit to maintain clarity and speed.