The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied plans and ideas, and it uses tools to find external facts, which creates an ingestion surface for potentially untrusted data. However, the skill includes a strong security control by mandating that the agent must wait for a "shared understanding" confirmation from the user before taking any action.
Ingestion points: User-provided plans and ideas in the prompt, and external facts found via "available tools" (SKILL.md).
Boundary markers: The skill requires the agent to internalize a summary of shared understanding and explicitly wait for user confirmation before proceeding.
Capability inventory: Fact-finding via tools; no high-privilege capabilities like file-writing or network exfiltration are requested in the skill definition.
Sanitization: Relies on a human-in-the-loop verification process as a functional safety barrier.
[REMOTE_CODE_EXECUTION]: There is no evidence of remote script downloading, package installation, or arbitrary command execution.
[DATA_EXFILTRATION]: The skill does not access sensitive system files or credentials, nor does it include instructions to send data to external servers.
[OBFUSCATION]: The instructions and evaluation tests are written in clear, readable Korean and English with no hidden characters, encoding, or homoglyph attacks.
[SAFE]: The skill implements a defensive design by ensuring that it does not act autonomously without user verification.

vibe-grilling