The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external data from GitHub and issue trackers, which could contain malicious instructions. \n- Ingestion points: The skill ingests data from GitHub CLI (gh), Git, and external artifact URLs via WebFetch (SKILL.md, swe.md). \n- Boundary markers: Instructions explicitly state that all fetched data must be treated as untrusted text to be summarized and never followed as instructions, with a requirement to flag potential injections (SKILL.md). \n- Capability inventory: The skill uses the Artifact tool for publishing, run_command for CLI access, and file tools like write_to_file for local state persistence (SKILL.md). \n- Sanitization: The skill implements mandatory HTML entity encoding for external content and JSON Unicode escaping to prevent script injection (SKILL.md, template.html). \n- [COMMAND_EXECUTION]: The skill executes development-related commands to gather project state. \n- Evidence: The skill uses gh and git via run_command to retrieve pull request details, CI status, and commit history (swe.md).

project-artifact