The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from local skill files and web search results to generate management reports and recommendations.
Ingestion points: The scripts/inventory_skills.py script reads the SKILL.md files of all locally installed skills to generate an inventory. The 'Explore' mode uses the websearch service to gather arbitrary content from external sources like GitHub and Reddit.
Boundary markers: While formal prompt boundary markers are not implemented for ingested data, the skill instructions repeatedly emphasize that both the AI and the user must manually read and evaluate any code or documentation before proceeding with installations.
Capability inventory: The skill can write to local skill files via scripts/inject_family_header.py, generate new files from internal templates, and perform network operations via git and websearch tools.
Sanitization: The skill includes a dedicated security triage script (scripts/scan_skill_security.py) that uses regular expressions to detect common malicious patterns (such as obfuscation, credential theft, and shell pipes) in third-party code before it is promoted from staging.
[COMMAND_EXECUTION]: The skill executes local Python scripts and the git command-line tool to perform inventory scans, modify skill headers, and clone repositories during the audit and installation workflows. These operations are restricted to user-owned skill directories as a safety measure.
[EXTERNAL_DOWNLOADS]: In 'Explore' mode, the skill researches and downloads third-party skills and plugins from external sources, primarily GitHub, for the purpose of staging and manual security evaluation.

skill-explorer