The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill executes code from a computed relative path outside of its own package structure.
The script scripts/run-audit.mjs calculates the location of a runtime script using new URL("../../../runtime/visual-audit/cli.cjs", import.meta.url) and executes it using node. This behavior depends on the skill being installed in a specific environment layout and executes code that is not contained within the skill's source.
[COMMAND_EXECUTION]: The skill launches an external process to perform the audit.
In scripts/run-audit.mjs, child_process.spawn is used to run the audit runtime, forwarding all command-line arguments (process.argv.slice(2)) to the child process. This allows for the execution of arbitrary flags or commands supported by the external runtime.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection as it ingests untrusted data from external sources and uses it to drive agent actions.
Ingestion points: Data is retrieved from the Figma API (figma-capture.md) and the browser DOM (browser-capture.md). The measurement script specifically extracts textContent and styles from arbitrary web pages.
Boundary markers: There are no delimiters or "ignore instructions" warnings defined for the data extracted from the browser or Figma.
Capability inventory: The skill has the ability to execute shell commands via run-audit.mjs and write files to the local .primeui/temp/ directory as part of the correction-loop.md.
Sanitization: The skill lacks explicit sanitization or filtering of the text content ingested from the browser or Figma. If a web page or Figma node contains malicious instructions, they could influence the agent during the "Correction Loop" phase where the agent is instructed to read the audit violations and apply changes to the codebase.

prime-visual-parity