Choose Claude Code Skills or Subagents for a Release Review
A skill can supply a standard procedure to the current conversation. A subagent can work through a focused assignment in its own context and return a result. Claude Code also supports running a skill in a subagent when the procedure needs its own context.
Worked example
Repeatable release check for the current conversation
--- name: release-check description: Review a draft release note for evidence and ordering. disable-model-invocation: true --- List the user-visible changes, cite the supplied release evidence, and flag uncertain claims for the author. Return at most five bullets.
Steps
- Decide what the current conversation needs to see. Save the sample checklist in .claude/skills/release-check/SKILL.md and invoke it in a release drafting session.
- For a long evidence search, give a read-only reviewer one folder, one release identifier, and a response format. Keep the returned evidence separate from the release text until a reviewer checks it.
- Try both approaches on a tiny release draft. Pick the arrangement that keeps the source references reviewable and exposes enough detail in the parent conversation.
What to check
- Check skill invocation by running /release-check and supply the draft and supporting files. A subagent's returned summary can leave out evidence that its parent needs.
- Use the skill's context setting only if you want it to run in a separate agent context. Keep human approval for publication outside the review step.