The Agent Skills Directory

Handling of External Data: The skill processes content from the repository (TODOs, FIXME comments) and GitHub issues. It treats all such external content as untrusted data and includes explicit instructions to ignore any embedded directives. This design helps mitigate potential indirect prompt injection risks by ensuring the agent does not follow commands found within the codebase or issues.
Subagent Validation Contract: When utilizing the repo-scout subagent, the skill enforces a strict output contract. Every value returned (counts, paths, IDs, links) must match specific regular expressions. For instance, paths are limited to 120 characters and must exist in the repo, while links are restricted to specific GitHub structures. If any returned value fails validation, the entire data set is discarded.
Safe Command Execution: The skill uses standard tools like gh (GitHub CLI), mktemp, and rmdir for administrative tasks such as exporting issues to a temporary directory. It explicitly forbids incorporating untrusted data into shell commands and ensures that temporary files are deleted after use. These operations are scoped to the intended functionality and include user consent checkpoints.
Privacy and Credential Protections: Instructions specifically prohibit copying secrets or credential files. It also advises users to keep sensitive data out of prompts during creative sessions, aligning with best practices for secure AI interactions.

hackathon