Liaison · Supervisor · WorkersLiaison · Supervisor · Workers

This is a Claude Code skill I wrote for running a project as a small AI team. I only ever talk to one window, and a team of Claude sessions and subagents does the work behind it. It runs as two Claude Code sessions in two terminals on the same project, and the skill itself is written in Chinese.

The liaison is the session I talk to: it relays my instructions, translates replies into plain language, checks facts and records decisions, and never writes code or touches API keys. The supervisor writes the code, makes the technical calls, splits the work and dispatches workers. Workers are subagents that each change only their assigned files, with the model picked by difficulty: a small one for small tasks, the strongest only for hard ones.

Setup takes one sentence in a project folder: the skill creates the shared files from its templates, including the liaison's rules, a progress log and a to-do list for me. Terminal 1 becomes the supervisor, terminal 2 the liaison, and from then on I only talk to the liaison. A big job can go to an optional manager subagent, which directs its own workers.