The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a surface for indirect prompt injection by requesting raw, unedited writing samples (Q2) and business descriptions from the user. This data is then used to populate the agent's core context files, which guide the agent's voice and decision-making in subsequent tasks.
Ingestion points: User-provided text for questions Q1 through Q7, initially stored in aios-intake.md and then processed into context files.
Boundary markers: The instructions do not define delimiters or structural boundaries when migrating user answers into the context/ and references/ files, making it difficult for the agent to distinguish between valid data and embedded malicious instructions.
Capability inventory: The skill has file system write capabilities, creating and updating files in the context/, references/, and archives/ directories, and modifying the primary operating manuals (AGENTS.md and CLAUDE.md).
Sanitization: The skill lacks sanitization, explicitly instructing the agent to paste voice samples 'verbatim' and 'raw', ensuring that any potentially malicious content in the user's writing sample is preserved in the agent's context.

onboard