The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses several command-line tools including gh (GitHub CLI) for repository scanning and bash for running a local linter (skill-lint.js). These operations are restricted to scanning repository structures and validating the format of generated files, with explicit rules in SKILL.md prohibiting the execution of any code or scripts found in external sources.
[EXTERNAL_DOWNLOADS]: The skill fetches repository metadata and file content from GitHub using the official GitHub API (gh api). It implements a Security Envelope that restricts fetches to HTTPS-only, enforces a 500KB payload limit, and uses a validateSecureUrl function to deny private, reserved, or link-local IP addresses (e.g., 127.0.0.1, 169.254.x.x).
[PROMPT_INJECTION]: The skill explicitly defines an 'Untrusted content rule' in SKILL.md and Phase 2 analysis. It treats all fetched content as data to be sanitized and analyzed, rather than instructions to be followed, which mitigates risks from instructions embedded in external skills or articles.
[DATA_EXFILTRATION]: While the skill reads external data, it does not demonstrate patterns of exfiltrating sensitive local data. Inbound data is subject to strict validation and sanitization before being processed by the LLM for skill generation.
[COMMAND_EXECUTION]: Path traversal is mitigated in Phase 0 argument validation using fs.realpathSync and path.relative to ensure that target directories for generated files remain within the designated project root.

sharingan