Some skills should run in isolation from your main conversation. Use context: fork in the frontmatter.
With context: fork, Claude spins up a separate subagent to run the skill. The subagent has its own context window, and when the skill finishes, the results come back to your main session.
Pair this with agent: Explore for read-only investigation tasks. The Explore subagent reads files and reports findings without modifying anything. Use agent: Plan when you want Claude to map out a plan before acting.
I find fork isolation most useful for skills that do broad research across many files, since it keeps all those reads out of your main context.