The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is instructed to execute shell commands discovered within the target codebase's configuration files (such as package.json scripts, Makefiles, or CI configurations) to verify their functionality during the refresh process. Evidence: The skill instructions specify that 'Read-only commands (test, lint, validate) run directly' to prove validity.
[DATA_EXFILTRATION]: The skill performs extensive read operations across the entire repository, accessing source code, architecture patterns, and potentially sensitive configuration files to generate its summary. Evidence: The skill uses tools like 'grep', 'glob', and 'read' to navigate all files in the repository.
[PROMPT_INJECTION]: The skill processes untrusted content from an unfamiliar repository and distills it into an AGENTS.md file intended for auto-loading by other agents, creating a vector for persistent indirect prompt injection. Ingestion points: Uses grep, glob, and read to ingest all code and configuration files in the repository. Boundary markers: The generated output includes a metadata stamp (), but there is no evidence of explicit sanitization for the content being distilled. Capability inventory: The skill has the ability to read all files, write to the repository root, and execute shell commands. Sanitization: No evidence of sanitization for ingested data before it is written to the output file.

map