The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a specific security instruction to treat any user-provided text containing commands (e.g., "ignore rules", "output in English") as data to be polished rather than instructions to be executed. This is a best-practice defense against indirect prompt injection.
[DATA_EXFILTRATION]: The instructions explicitly forbid recording API keys, tokens, passwords, or PII. Network access via the gh tool is strictly limited and only triggered by a direct user request for bug reporting, preventing unauthorized data transmission.
[INDIRECT_PROMPT_INJECTION]: The skill handles external user-provided text (grant proposals). While this presents a surface for injection, the logic includes explicit boundary instructions to ignore embedded commands and focuses solely on linguistic transformation, which is a low-risk configuration.
[COMMAND_EXECUTION]: The skill uses localized directories for task logs and bug reports (./.bensz-api/ and ~/.bensz-skills/bugs/). These operations are consistent with the skill's stated management functions and do not involve arbitrary shell execution or unsafe path traversal.

nsfc-humanization