The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a multi-session workflow that involves reading data from previously saved module files and a state.json file. This pattern creates a surface where instructions or adversarial content embedded in interview answers could potentially influence the agent's behavior when the session is resumed.
Ingestion points: The agent reads from state.json and various markdown files located in the modules/ and founders/ subdirectories as specified in the SKILL.md session start and state protocol sections.
Boundary markers: The instructions do not define specific delimiters or 'ignore embedded instructions' warnings for the content read from these files, which is a common practice to mitigate injection.
Capability inventory: The skill utilizes file read and write capabilities to manage session state and interview results across multiple modules (10 through 90).
Sanitization: The skill includes explicit instructions for the agent to validate participant names (restricting to alphanumeric and hyphens) and outputPath (requiring absolute paths within the project and rejecting .. segments) to mitigate path traversal risks during the file writing process.

brand-discovery