The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to perform a directory listing (ls) to detect existing instruction files such as CLAUDE.md, GEMINI.md, and .github/copilot-instructions.md. This is a standard discovery operation to determine the project's environment.
[EXTERNAL_DOWNLOADS]: References a public GitHub repository (github.com/forrestchang/andrej-karpathy-skills) as the source for the behavioral principles it implements. This reference is used for attribution and documentation and does not involve the automated execution of remote code.
[PROMPT_INJECTION]: The skill analyzes and rewrites existing instruction files which may contain untrusted data from external repository contributors, representing an indirect prompt injection surface.
Ingestion points: Reads project-level documentation files (CLAUDE.md, GEMINI.md, etc.).
Boundary markers: While it does not use specific delimiters when reading input, it enforces a strict "Canonical Structure" during the rewrite phase.
Capability inventory: The skill possesses Read, Write, and Bash capabilities to manage repository configuration.
Sanitization: Implements a structured "Quality Checklist" designed to identify and remove "stale artifacts" or "foreign rules" from other projects, which serves to sanitize the resulting instructions.

optimize-agent-instructions