The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches prompt engineering guidelines from official sources (platform.claude.com and claude.com). These are well-known, trusted domains for the intended purpose of the skill and do not present a security risk.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided prompt ideas and existing prompts. While this is an ingestion surface for untrusted data, the skill uses structured templates, explicit boundary markers (e.g., <context>, <task>, <output_format>), and includes a mandatory 'Test Run' and grading phase to verify output safety and efficacy. These measures significantly reduce the risk of injection affecting the agent's core logic.
[COMMAND_EXECUTION]: The skill includes instructions for running test prompts within 'Claude Code' subagents. This is a platform-native feature for testing prompts in isolated environments and does not involve arbitrary or dangerous shell command execution on the host system.

prompt-lab