The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's audit script (scripts/names.py) ingests data from external SKILL.md files to detect name collisions.
Ingestion points: The script reads SKILL.md files from ~/.agents/skills/ and project-local .agents/skills/ directories.
Boundary markers: No specific delimiters or safety instructions are used when processing the content of these manifests.
Capability inventory: The skill instructions involve executing shell commands such as npx, git, and python3, providing a potential execution path for injected instructions.
Sanitization: The script performs hashing and basic name normalization but does not sanitize the markdown content against embedded instructions.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to interact with external registries and repositories.
Registry Query: scripts/names.py performs HTTP GET requests to https://skills.sh/api/search to check for skill name availability.
Package Execution: The instructions call for using npx to execute the skills CLI, which may download the package if not present.
Repository Cloning: The skill documents the use of skills add <owner>/<repo>, which performs git clones of remote repositories.
[COMMAND_EXECUTION]: The skill relies on several CLI tools to perform its tasks.
Tooling: Instructions include the use of npx, git, python3, and symlinking (ln -s) to manage the skill environment.

skill-authoring-taudelta