Skip to main content
Skills and sub-agents solve problems that sound identical and are not. Both “make Claude better at a specialized task.” The difference is architectural, and once you see it you will never confuse them again: A skill is a playbook. It loads into the conversation you are already having, and the Claude in front of you follows it while you watch and steer. A sub-agent is a worker. It is a separate Claude with its own private context, briefed for one job. It goes away, does the work, and only its final summary comes back to you.

​
Run the same job both ways

Two jobs below, each runnable as a skill in your main thread or as a dispatched sub-agent. The columns are context windows. Watch what each strategy does to them.
The two jobs disagree on purpose. The 40-contract audit floods whatever context runs it, so it belongs in a worker whose memory you can throw away. The memo walkthrough needed a mid-course correction only you could give, so it belongs in front of you, guided by a playbook. Neither strategy is the better one; they are answers to different questions.

​
The three questions

When a specialized task shows up, ask in order:
  1. Will the working steps produce pages of output I will never reread? If yes, sub-agent. Intermediate noise is exactly what isolated context is for.
  2. Will I need to steer mid-task? If yes, skill in the main thread. A sub-agent cannot hear you until it is finished.
  3. Does the work need different permissions or a different model? Sub-agents can run with restricted tools, in parallel, or on a cheaper model. Skills always inherit the current conversation’s everything.
If none of the three fires, default to a skill. It is lighter to write, easier to debug, and keeps the work where you can see it.

​
Side by side

​
The combination is the real pattern

Mature setups stop choosing and start stacking: sub-agents that carry skills. The playbook teaches, the worker executes. In Claude Code, a sub-agent’s configuration can preload specific skills into its context at startup, so your contract-audit worker begins life already knowing the firm’s audit checklist. Teams converge on the same shape: a small bench of workers, three to six, and a deep shelf of playbooks, ten to thirty, shared across them. Carry the sorting instinct: knowledge becomes a skill, labor becomes a worker, and the best workers are handed your skills on their way out the door.

Comparison grounded in Anthropic’s guide to steering Claude Code and the sub-agents documentation.