Delegate a sub-task to a Cursor agent. Use when a task is better handled by a Cursor agent (e.g. complex file editing, multi-step coding tasks, research with web access).
Installs into .claude/skills of the current project.
Are you the author of Delegate?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/brainrot-creations-delegate-df42db51)
---
name: delegate
description: Delegate a sub-task to a Cursor agent. Use when a task is better handled by a Cursor agent (e.g. complex file editing, multi-step coding tasks, research with web access).
---
Use the `cursor-agents` MCP tools to break work into sub-tasks and delegate them to Cursor agents.
## Synchronous delegation (need result now)
Use `cursor_run` when you need the output before continuing:
```
cursor_run(
prompt="<full task description with all context the agent needs>",
cwd="/path/to/project" # optional, defaults to current directory
)
```
Returns `{ text: "<agent output>", stop_reason: "..." }`.
## Background delegation (parallel work)
Use `cursor_task_submit` to fire off tasks and continue working:
1. Submit one or more tasks:
```
cursor_task_submit(name="Refactor auth module", prompt="...", cwd="...")
cursor_task_submit(name="Write tests for API", prompt="...", cwd="...")
```
2. Check status:
```
cursor_task_status(task_id="<id>")
```
3. Fetch result when completed:
```
cursor_task_result(task_id="<id>")
```
4. See all tasks:
```
cursor_tasks_list()
```
## Writing good prompts for Cursor agents
- Include full context — the agent has no knowledge of what you've done so far
- Specify the working directory if the task involves files
- Be explicit about what "done" looks like (files to edit, tests to pass, output expected)
- If the task spans multiple files, list the key ones