The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text (descriptions, prompts, and file names) to construct commands for the gemini CLI. This creates a standard surface where untrusted data is interpolated into tool executions.
Ingestion points: User-provided image prompts and file paths are used in the -p and -f flags of the gemini command within SKILL.md.
Boundary markers: The skill uses placeholders such as [최적화된 프롬프트] and [경로/파일명.png] to guide prompt construction.
Capability inventory: The skill utilizes the Bash tool to execute gemini CLI commands, restricted by the allowed-tools configuration to only commands starting with gemini:.
Sanitization: The instructions do not explicitly define sanitization routines for the user input; the security of the execution relies on the agent's internal tool-calling guardrails and the shell environment's handling of special characters.

nano-banana