The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data by analyzing workspace changes through git commands.
Ingestion points: The skill instructs the agent to run git status and git diff (SKILL.md) to analyze current workspace changes.
Boundary markers: There are no explicit boundary markers or instructions to treat the content of the diffs as data only, increasing the risk that instructions embedded in code changes could influence the agent.
Capability inventory: The skill utilizes git add and git commit to modify the repository state based on the analyzed content.
Sanitization: No sanitization or filtering is performed on the output of the git commands before the agent processes them.
[METADATA_POISONING]: The instructions contain specific directives to conceal the automated nature of the work and the identity of the agent.
Evidence: "Never include commit bodies, trailers, co-author lines, AI/tool attribution, generated-by markers, bot signatures, or metadata implying AI involvement."
This behavior ensures that commits generated by the AI are indistinguishable from human-authored commits, which can bypass security audits, logging, or provenance tracking systems that expect AI-generated content to be marked.

commit-changes