Les primitives Claude Code: qui invoque quoi
Command, skill, sub-agent, agent team, loop, goal, MCP: the building blocks of Claude Code look alike enough to be mixed up, and the mix-up gets expensive when you hand one the job of another. Two questions sort them out. First: who triggers it? A command is you invoking it (you type the slash). A skill is Claude reaching for it when the task fits. An MCP plugs Claude into outside tools (calendar, email, files); a skill adds know-how, not tools. Second: when does it stop? A loop reruns a prompt on a clock, again and again. A goal works until a condition is met, then stops on its own. A sub-agent works in isolation and returns a summary; an agent team shares a common report and coordinates.
Strengths
- Two questions are enough to decide between seven look-alike building blocks
- Avoids the loop/goal trap that runs an agent in circles or cuts it off too early
- No code needed: a PM or designer can pick the right primitive on the first try
Limitations
- The grid sorts the primitives but doesn't say how to write them: for a skill or agent you then move to the file format
- The Claude Code vocabulary moves fast: a name or behavior can shift from one version to the next
Best for
- A PM or PO discovering Claude Code who isn't sure whether they need a command, a skill or an agent
- A designer wanting to automate a repetitive task and torn between a loop and a goal
- A DS-manager training a team on the tool and looking for a simple grid to pass on