The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides extensive instructions for executing system commands across macOS, Linux, and Windows to perform screen captures and document conversions (e.g., google-chrome, screencapture, pdftoppm). This includes a PowerShell snippet in references/auto-evaluation-visuelle.md that uses Add-Type to dynamically load .NET assemblies (System.Windows.Forms and System.Drawing) to capture the virtual screen.
[EXTERNAL_DOWNLOADS]: In references/auto-evaluation-visuelle.md, the skill suggests using npx playwright screenshot, which involves fetching and executing the Playwright tool from the npm registry at runtime.
[DATA_EXFILTRATION]: The screen capture instructions, particularly native OS commands like screencapture -x or the PowerShell virtual screen capture, have the potential to capture the user's entire desktop or sensitive windows if the agent is not strictly constrained to the intended artifact.
[PROMPT_INJECTION]: The skill processes untrusted artifacts such as HTML pages and PDFs, which provides a surface for indirect prompt injection that could influence agent behavior during audits or visual evaluations.
Ingestion points: Local files and absolute paths opened via browsers or document tools as described in references/auto-evaluation-visuelle.md.
Boundary markers: None identified in the instructional templates.
Capability inventory: Shell and PowerShell execution, headless browser navigation, and file system access across reference files.
Sanitization: No sanitization or content validation of the ingested artifacts is specified.

fabuleux