The Agent Skills Directory

[PROMPT_INJECTION]: The skill identifies and attempts to mitigate indirect prompt injection risks arising from its primary function of processing untrusted external data.
Ingestion points: The skill reads item titles, bodies, comments, and pull request text from external issue trackers using the gh CLI.
Boundary markers: It references an item-content-trust.md guide and explicitly instructs the agent to treat all item text as data to be evaluated, never as instructions to be obeyed.
Capability inventory: The skill has the capability to execute shell commands (gh, git), write to the local file system, and schedule future cycles.
Sanitization: The instructions explicitly require using JSON parsers for untrusted fields and prohibit the string interpolation of user-influenced data into shell commands or prompts.
[COMMAND_EXECUTION]: The skill uses shell scripts to perform telemetry updates and repository maintenance.
It executes gh api commands to search for, create, and update telemetry comments on GitHub issues.
It uses git commands (e.g., git check-ignore, git rev-parse) to manage local repository configuration, such as ensuring escalation records are ignored by version control.

work-loop