Cursor CLI Delegation

Delegate the requested work through Cursor's own CLI and return the requested answer or artifact with supporting evidence. Cursor can own planning, experiments, execution, and verification; an already-settled plan is not a prerequisite. Choose the division of work for the task instead of imposing a caller-supervisor role.

Defaults And Choices

• Launch only when the user explicitly invokes this skill for delegation. Discussing or editing it does not authorize a model call.
• Use the user's exact parent model when supplied, verified against the live model listing; otherwise omit --model so the account's default model runs. Do not guess an alias or silently substitute a model.
• Choose mode and approval by the task's authority. Work that carries no edit authority, such as planning, research, review, or answering questions, runs in native read-only mode (--mode plan for planning, --mode ask otherwise) without YOLO. Use general agent mode with YOLO only for granted edit work, including commands the user asked Cursor to run, subject to narrower user or host policy. YOLO is broad unattended approval, not additional task authority. Preserve sandbox configuration.
• An explicit user request for a mode overrides this default when the task's authority allows it; producing a plan alongside granted edits does not itself require plan mode.
• Resume the exact returned session for related follow-ups, refinements, and continued work. Start fresh for independent work.
• Let Cursor choose useful internal subagents and their models unless the user supplies a child-model policy. Parallel work is optional; sequential execution remains valid.

Working Path