Every sprint you paste the same five-step release procedure into the conversation. Two of its steps need subagents spawned and several files read, and the procedure runs in the conversation you are already in. Where should it live?

The principle — the question is not what the procedure does, it is where a procedure you keep repeating should live.

Why the key is correct — Anthropic names the trigger for creating a skill as exactly the situation described: you keep pasting the same instructions, checklist, or multi-step procedure into chat. The second half of the question is the part people get wrong, and the documentation settles that too — bundled skills give Claude a detailed prompt and let it orchestrate the work, so they can spawn agents, read files, and adapt to your codebase. A skill is not restricted to reciting text.

Why the others are wrong — keeping the agent-spawning steps manual gives up the property that made a skill fit in the first place. A subagent is a different thing: skills are what to consider when the workflow should run in the main conversation context rather than in isolated subagent context, which is where this procedure runs. A dynamic workflow is a real option in general, but not for this reason — it is for work needing more agents than one conversation can coordinate, or for orchestration you want codified as a script you can read and rerun, and spawning two workers clears neither bar.

Remember this — a procedure you keep pasting in belongs in a skill, and a skill can orchestrate rather than merely recite.

Sources — Anthropic's Claude Code skills, glossary, subagents and dynamic workflows documentation.