An unofficial reference for people writing Claude skills. Every claim here is derived from shipped binaries and live sessions, and carries the build it was checked against and how confident that check was.
Claude Cowork runs your skill in a sandbox whose rules differ from the Claude Code CLI in ways that silently break otherwise-correct skills — files that never reach the user, deletes that fail, a shell that cannot see your environment. This documents those differences.
What your skill can rely on inside Claude Cowork — paths, file delivery, deletes, the shell, runtime detection, plugins and sub-agents.
Verified 2026-08-05
Not official, and not affiliated with Anthropic. It is reverse-engineered from shipped software. It does not detect product changes — behaviour served from Anthropic's side can change at any time, with no version bump and no signal here. Claims are stamped with a date and a confidence label so you can judge how much weight to give them.