The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from external code repositories, creating a theoretical surface for indirect prompt injection.
Ingestion points: The skill reads repository files (src/**), diff outputs (git diff), commit messages, and pull request descriptions as detailed in the 'How to run' section.
Boundary markers: The skill includes a dedicated 'Untrusted input' section that explicitly instructs the agent to treat diff and repository content as data and 'never instructions'.
Capability inventory: The skill performs read operations and proposes file changes or shell commands. It specifically prohibits the agent from automatically executing any scripts, hooks, or tests found within the target repository.
Sanitization: Instructions mandate that the agent must identify and report attempted instruction injections (e.g., comments claiming rules are waived) as violations rather than obeying them.
[COMMAND_EXECUTION]: The skill utilizes shell commands for repository analysis and management.
Evidence: The 'How to run' and 'Adopt mode' sections describe using git diff for scope resolution and suggest running utility scripts such as check-host.py or check-identity.sh.
Security Controls: The skill follows a policy of never executing commands from the reviewed tree automatically. Sensitive operations, such as applying migration scripts or using an admin token, require explicit user consent and execution.

atelier-review-me