The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is configured to load and prioritize instructions from .github/aw/instructions.md if the file exists within the local repository. This represents a potential surface for indirect prompt injection, as content within the repository can influence the agent's behavior for any user invoking the skill in that context.
External Resource Retrieval: The skill fetches extensive workflow logic, templates, and instruction sets from an external repository (github/gh-aw). This is a core part of its design as a central dispatcher for workflow management across different task types.
Command Execution Best Practices: The included documentation in references/runtime-contracts.md provides detailed security guidance, specifically advising against the interpolation of event data directly into shell scripts. It recommends using environment variables and quoting values to mitigate potential command injection risks.
Credential and Privilege Management: The skill emphasizes maintaining strict security boundaries between execution phases (e.g., trusted pre-agent work vs. agent sandbox). It cautions against widening token permissions or exporting write-privileged credentials into the sandbox environment, advocating for the principle of least privilege.

agentic-workflows