The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests data from a jtbd.json corpus to guide the design process, creating an attack surface for indirect prompt injection. Malicious instructions within the corpus could subvert the agent's behavior. \n
Ingestion points: jtbd.json corpus (referenced in SKILL.md). \n
Boundary markers: The skill does not define explicit boundary markers or instructions to ignore embedded instructions within the corpus data. \n
Capability inventory: Local file writing (via artifact_path) and design document modification (via backend tools). \n
Sanitization: No sanitization, validation, or escaping of the corpus content is described. \n- [COMMAND_EXECUTION]: The skill utilizes external design tools (such as Pencil) that exhibit unsafe behavior by ignoring provided file paths and writing directly to the active document. This presents a risk of unintended data modification or overwriting if the agent fails to strictly enforce user confirmation steps. \n
Evidence: The references/design-file.md documentation notes that the backend tools ignore the filePath argument and overwrite the active document regardless of the intended target. \n- [DATA_EXFILTRATION]: The skill enables the creation and potential publication of artifacts. If the agent's generation logic is subverted by malicious input (e.g., via the task corpus), these capabilities could be misused to exfiltrate data from the project or distribute malicious HTML/SVG files containing unintended scripts or content.

prototype