The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external websites, repositories, and local files during Phase 1. This content is processed to build a 'knowledge map', creating a surface where malicious instructions embedded in source material could potentially influence agent behavior.
Ingestion points: Website URLs (via browser/fetch tools), Repositories (README, docs, code), and local files/folders.
Boundary markers: The skill explicitly instructs the agent to treat fetched data as evidence only and never as instructions, specifically stating that source text cannot change the campaign's parameters or authorize network/file operations.
Capability inventory: The skill can write files to the local disk, perform network research, and execute shell commands (explorer, open, xdg-open, magick, pandoc).
Sanitization: Facts are distilled into a structured 'knowledge map' and the final output undergoes a multi-stage self-audit process (Phase 6) to detect and remove sensitive data or machine-like patterns.
[COMMAND_EXECUTION]: The skill invokes local system utilities to perform specific tasks.
It uses explorer (Windows), open (macOS), or xdg-open (Linux) to open the generated campaign folder for the user.
It checks for the presence of and executes magick (ImageMagick) for image format conversion.
It checks for pandoc to convert Markdown files to PDF.
It utilizes system video tooling (e.g., ffmpeg) to process supplied video files into stills and shorts.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to gather information.
It fetches content from user-provided URLs and clones remote repositories during the ingestion phase.
It performs live web research in Phase 3 to verify platform-specific constraints, rules, and engagement studies.

awesome-content-campaign