Skill vs. (Sub)Agent

New series! From now on, I'll regularly share short lessons from our day-to-day work with Claude Code. Let's kick off with a question I kept asking myself early on:
"Should I build a skill or a subagent for this?"
Short answer: they solve different problems.

🧠 Skill = knowledge and workflows on demand

A skill is essentially a Markdown file (SKILL.md), optionally bundled with scripts and templates. Claude loads it automatically when it matches the task, or I invoke it directly via /command. It runs in the current session and shares its context.
Typical uses: style guides, deployment checklists, recurring workflows.

🤖 Subagent = a team member with their own desk

A subagent works in a fresh, isolated context with its own system prompt, tools and permissions. It handles one well-defined task and only reports back the result. Your main session stays lean and focused.
Typical uses: code reviews, security checks, searching large codebases.

My rule of thumb:

• Claude needs to KNOW something or follow a fixed PROCESS? Skill.
• A task should run in ISOLATION without flooding the main context? Subagent.

The best part: it's not either/or. A subagent can preload skills, e.g. a review agent that brings our coding style guide along.

How are you using skills and subagents? I'd love to hear your experiences in the comments

Join the discussion on LinkedIn →