Skills and sub-agents
A capable generalist becomes an expert team through two mechanisms : procedures and specialists. Claude Code has both, under the names skills and sub-agents.
Skills : recipe cards for the agent
A skill is a written procedure : a text file describing, step by step, how to perform one recurring task the way you want it done. When the task comes up, the agent pulls out the recipe and follows it : exactly, every time.
The kitchen analogy holds well : a good cook can improvise a dish ( the model's general ability ), but a restaurant runs on recipes so that the dish is identical whoever is in the kitchen tonight. Skills are those recipes.
Examples of real skills from setups we operate :
- « Deploy the website » : the exact sequence of checks and commands, including the verification that every page actually responds afterwards.
- « Monthly report » : where the data lives, which numbers to compute, the exact format, who receives it.
- « Before believing any statistic » : a checklist of statistical pitfalls to check first ( yes, a skill can encode judgment, not just steps ).
The power move : when the agent does something well after a few iterations, turn that into a skill. Your know-how stops living in people's heads and starts living in versioned files : readable, improvable, transferable.
Sub-agents : specialists you can dispatch
Sometimes one collaborator is not enough. A sub-agent is a separate agent that the main agent can dispatch for a piece of work : with its own focus, its own instructions, and its own working memory ( so it does not clutter the main conversation ).
Three situations where this shines :
- Parallel work : eight specialists each audit one aspect of a website simultaneously : technical, content, performance... : and the main agent assembles their findings. What took a day takes twenty minutes.
- Specialization : a « reviewer » sub-agent with strict instructions checks the work of the main agent before anything ships. Producer and inspector are kept separate : just as in a well-run team.
- Big reading tasks : a sub-agent reads a 200-page document and returns the two paragraphs that matter, keeping the main conversation clean.
If skills are recipes, sub-agents are colleagues with a job description. You write the job description once ( « you are a meticulous reviewer ; refuse anything without proof » ) and dispatch them whenever the situation calls for it.
How skills and sub-agents combine
The patterns from the agentic workflows page come alive here : an orchestrator ( the main agent ) routes work to specialists ( sub-agents ), each following its procedures ( skills ), inside the guardrails ( permissions and hooks ). That sentence is essentially the architecture of every serious agentic system we have built at eaQbe.
Frequently asked questions
What is the difference between a skill and just asking the agent?
Repeatability. Asking relies on remembering to ask correctly each time ; a skill is the written procedure, triggered when relevant, identical every time, and improvable in one place.
When should I create a skill?
The third time you explain the same procedure : write it down as a skill. As a rule of thumb : recurring + has a « right way to do it » = skill material.
What is a sub-agent in simple terms?
A second AI worker, with its own instructions and its own focus, that the main agent can send on a mission : reviewing, exploring, or handling one part of a bigger job in parallel.
Going further
- Next : MCP & persistent memory : connecting the agent to your real systems.
- In the Claude Code training, you write your first skill on a process from your own work.