The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted source code files (.jsx, .tsx, .html, etc.) from the user's project and external website content via Lighthouse to generate accessibility reports. This creates a surface where instructions hidden in code comments or website metadata could potentially influence the agent's behavior during the reporting phase.
Ingestion points: Project source files are read via grep and eslint. External website content is processed via npx lighthouse (as seen in SKILL.md).
Boundary markers: Absent; there are no specific instructions to the agent to disregard or isolate embedded instructions within the content being audited.
Capability inventory: The skill can execute shell commands (npx, grep), write audit results to the project directory, and interpret findings to provide remediation advice.
Sanitization: Absent; the content is processed directly for patterns and linting rules without a sanitization layer.
[COMMAND_EXECUTION]: The skill instructs the agent to execute several shell commands, including complex grep pipelines, npx eslint, and npx lighthouse. These commands are used legitimately for static analysis and performance auditing within the scope of the skill's purpose.
[EXTERNAL_DOWNLOADS]: The skill utilizes npx to run eslint and lighthouse, which may download packages from the npm registry. This targets a well-known and established service (npmjs.com) and is documented here as part of the tool's standard operation.

web-accessibility-audit