The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface via image processing.
The derive_outline function in kernel.py uses a vision model to extract instructions and metadata (such as panel messages and 'asks') from user-provided PNG files. Maliciously crafted images containing hidden text instructions could attempt to influence the agent's behavior or output.
The skill documentation includes an explicit security control, instructing the agent that the image is "untrusted input" and that all derived strings must be reviewed and edited by the agent before proceeding to the next step.
[COMMAND_EXECUTION]: Execution of image processing and sub-agent tasks.
The skill utilizes PIL (Pillow) and matplotlib within the Python REPL environment to render, crop, and compose images. These operations are core to the stated purpose of the skill.
The skill uses host.delegate to fan out tasks to sub-agents for individual panel generation. This represents an architectural choice for parallelizing the figure composition workflow and is handled within the platform's execution boundaries.

figure-composer