The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-generated content from GitHub Pull Requests and Issues, which presents a surface for indirect prompt injection. However, the skill provides specific instructions to the agent to treat this data as non-authoritative and to disregard any embedded commands or behavioral overrides.
Ingestion points: Tools like gh-llm pr view and gh-llm issue view (in SKILL.md) read external timeline and comment data.
Boundary markers: A dedicated "Security: handling third-party content" section explicitly warns the agent against following instructions embedded in third-party text.
Capability inventory: The skill utilizes tools for writing comments, submitting reviews, and merging PRs, as well as local file writing for body templates.
Sanitization: The skill employs explicit instructional guardrails to prevent the agent from executing content found in GitHub data.
[EXTERNAL_DOWNLOADS]: The instructions describe the installation of the gh-llm tool and the ShigureLab/gh-llm extension, which are resources provided by the skill's author.

github-conversation