A lightweight workflow discipline that forces Claude to get explicit approval before executing anything with side effects. You create a packet with intent and scope, review it to APPROVED or REJECTED status, then dispatch it for execution. The clever bit is scope lock: once approved, the packet's intent is frozen, so if the task grows mid-flight you dispatch what you have and create a new packet for the rest. This directly attacks the context bloat problem where a simple task balloons into a session-killing monster. If you're running Claude in any kind of automated loop or delegating real work, this gives you a paper trail and a forcing function to keep tasks atomic.
npx -y skills add open-gitagent/opengap --skill packet-factory --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
juliusbrussee/caveman
mattpocock/skills
shadcn/improve
obra/superpowers
forrestchang/andrej-karpathy-skills
vercel-labs/skills