Your team's /map-dependencies skill reads dozens of files and produces several thousand lines of intermediate analysis before reaching a one-page conclusion. Developers run it mid-task, and afterwards the main session is cluttered and Claude loses track of the task they were working on. What is the most effective change to the skill?

Answer: B — Add context: fork to the skill's frontmatter so it runs in an isolated sub-agent context and only its final result returns to the main conversation, leaving the session uncluttered.

context: fork runs the skill in an isolated sub-agent context so its verbose exploration doesn't pollute the main conversation, which fits a skill with long internal work and short useful output (task 3.2). Asking for brevity (a) works against the analysis the skill exists to do and is unreliable. allowed-tools (c) controls which tools the skill can use, not how much output reaches the main context. /compact after every run (d) is a lossy manual cleanup step that treats the symptom each time.