The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of data from external websites and social media platforms (LinkedIn, Twitter, Facebook) through the use of browser automation tools for auditing and research purposes. This ingestion of untrusted data creates a vulnerability to indirect prompt injection.
Ingestion points: Detailed in references/shared-patterns.md, which provides workflows for fetching text and HTML from live URLs.
Boundary markers: The instructions do not define specific delimiters or 'ignore' directives to prevent the agent from following instructions found in external data.
Capability inventory: The skill has permission to save files to the project directory, including scripts, calendars, and performance reports, as defined in references/deliverables.md.
Sanitization: No specific sanitization or validation logic is described for the data retrieved from external sources.
[EXTERNAL_DOWNLOADS]: The skill setup instructions suggest downloading and installing a browser automation tool from a third-party source.
Evidence: references/shared-patterns.md references the installation of agent-browser from the vercel-labs GitHub organization.
Source: The download source is a well-known and established technology organization.
[COMMAND_EXECUTION]: The skill's instructions involve executing shell scripts and CLI commands to manage the environment and perform research.
Evidence: The skill references execution of tool-discovery.sh and chrome-profiles.sh within a sibling skill directory, as well as agent-browser CLI operations in references/shared-patterns.md.

paw-mkt-video