soplisted

# Standard Operating Procedure (SOP) ## What this is SOP is the AI's own working method — the default way it moves a work item from "accepted" to "ratified by a human," filling the delegated region between the human's intent and the human's final decision (see the three points in the **mission-command** skill). Where **roe** is what the human forbids, SOP is what the AI does by default when no special instruction says otherwise. The AI owns and may improve its own SOP; it may not touch ROE (see the **roe** skill's authoring rule). ## The default procedure 1. **Make progress visible first.** The very first action on picking up a work item is to mark it as in-progress (so parallel agents — human or AI — don't double-start the same item) and post a short start note: who/what, which branch, the intended scope, and how to resume if the work is interrupted. Only after that, plan. 2. **Isolate the work.** Use a dedicated branch (and, where supported, an isolated working copy such as a worktree) so parallel work items don't collide. Install dependencies before running anything so failures are attributable to the change, not the environment. 3. **Delegate implementation to subagents.** Don't implement directly in the orchestrating context — dispatch the work. See the **subagent-driven-development** skill for the mechanics and model-routing. 4. **Use advance-approval gates for low-reversibility or subjective areas.** Get human sign-off on a mockup or plan _before_ building, not aft