The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes explicit instructions in Phase 4 intended for a secondary reviewer model that command it to ignore standard safety and defensive behaviors. Specifically, the prompt includes directives such as 'No over-defense,' 'No defensive scaffolding,' and 'no caveat sprinkling' to force blunt feedback, which functions as an instruction override for the sub-agent.
[COMMAND_EXECUTION]: The instructions in Phase 3 provide shell commands to be executed via the Bash tool. These commands launch a local browser (Microsoft Edge) in headless mode to capture screenshots of generated files and use --dump-dom to export SVG content to the local filesystem.
[DYNAMIC_EXECUTION]: The skill's primary architecture involves the generation of a p5.js-based web page (figure.html) that is subsequently executed in a browser environment. This runtime execution of dynamically generated code is used to produce the visual figures and SVG exports.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest an external [composition source]—a URL or local path to an existing figure—which is then processed by a vision-capable reviewer. This ingestion of untrusted visual data represents an attack surface where malicious patterns or instructions embedded within the source image could influence the agent's iterative design process.
Ingestion points: [composition source] provided in $ARGUMENTS and processed in Phase 4.
Boundary markers: The instructions tell the reviewer to 'Judge only from this image,' but there are no technical sanitization steps for the visual input itself.
Capability inventory: Uses Bash to run browsers, Write/Edit to modify code, and WebFetch to retrieve data.
Sanitization: No visual sanitization or instruction filtering is applied to the input source image.

method-figure-loop