The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a delegation pattern where instructions ('briefs') provided by a user or orchestrator are passed to an external agent with file-system and command-execution capabilities.
Ingestion points: The skill reads tasks from brief.txt or stdin (detailed in references/dispatch-and-poll.md).
Boundary markers: Instructions are structured using XML-style tags (e.g., <task>, <action_safety>, <verification_loop>) to help the implementer distinguish the task from constraints (detailed in references/writing-the-brief.md).
Capability inventory: The delegated agent (cursor-agent) possesses extensive capabilities to modify the working tree and execute shell commands, such as tests, linters, and build scripts (detailed in SKILL.md).
Sanitization: The skill mandates that the orchestrator must perform a manual review of all diffs (touchedFiles) and re-run all project gates locally before committing any changes (detailed in references/review-and-land.md).
[COMMAND_EXECUTION]: The skill relies on the execution of the cursor-agent CLI to perform automated coding tasks. It explicitly documents the use of the --force flag for automatic command execution and the --trust flag to bypass workspace trust prompts, while advising that these should only be used in trusted repositories.
[EXTERNAL_DOWNLOADS]: The documentation directs users to install the Cursor CLI from cursor.com/cli, which is an official source for the service. It also references an external script (relay.mjs) that is not included in the skill bundle, explicitly advising users to obtain and review the upstream runtime separately.

cursor-delegate