The Agent Skills Directory

[COMMAND_EXECUTION]: Employs standard Git commands (git branch, git status, git log) to determine the repository's current state and suggest appropriate workflow steps. These commands are non-destructive and limited in scope.
[DATA_EXPOSURE]: Reads local repository metadata, including branch names and a limited set of recent commits and status entries. This information is processed locally to assist the user, with no evidence of exfiltration or unauthorized access.
[PROMPT_INJECTION]: Contains "Override boundary" instructions that explicitly direct the agent to adhere to the skill's defined workflow structure and resist potential redirection from external repository-level configurations.
[PROMPT_INJECTION]: Exhibits a surface for indirect prompt injection by reading repository data such as branch names and commit messages.
Ingestion points: Git command outputs and local project artifacts in the .work/ or memory directory.
Boundary markers: The skill requires verification of stages through specific artifacts and command outputs rather than conversation history.
Capability inventory: Read-only repository metadata access and local markdown file management.
Sanitization: The skill focuses on the presence of specific artifacts rather than executing instructions from parsed repository data, which mitigates the risk of indirect injection.

workflow