The Agent Skills Directory

[PROMPT_INJECTION]: The skill does not contain instructions that attempt to override the agent's core safety guidelines or operational constraints. The language instruction (defaulting to Russian) is a localization feature and does not constitute a jailbreak or safety bypass.
[DATA_EXFILTRATION]: There are no network operations, such as curl or fetch, and no hardcoded credentials or secrets. The skill reads project metadata (e.g., package.json) and documentation (e.g., README.md) to extract marketing information, but this data remains within the local project environment.
[COMMAND_EXECUTION]: The skill does not involve the execution of shell commands or subprocesses. Its primary functions are file system navigation (checking for existing context files) and text generation based on project content.
[REMOTE_CODE_EXECUTION]: The skill does not download or execute external scripts, nor does it define any package dependencies that could facilitate remote code execution.
[INDIRECT_PROMPT_INJECTION]: The 'auto-draft' feature reads untrusted data from the codebase (such as README files or marketing copy) to generate its output. While this is an ingestion surface for indirect injection, the skill lacks high-privilege capabilities (like network access or arbitrary command execution), which significantly limits the potential impact of such an attack.

product-marketing