The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, creating a potential surface for indirect prompt injection. 1. Ingestion points: The skill reads external content from GitHub issue titles and comments using gh issue view <n> --json title,comments as described in SKILL.md. 2. Boundary markers: No explicit delimiters or instructions to ignore embedded commands are specified for the ingested content. 3. Capability inventory: The skill can create, edit, and reopen GitHub issues, create labels, and perform Git operations (commit, push) via the CLI. It also has the ability to write to local files like CONTEXT.md and .scratchpad/. 4. Sanitization: There is no evidence of sanitization or filtering of the content retrieved from GitHub issues before it is processed by the agent.
[DYNAMIC_EXECUTION]: The instructions suggest that the agent should generate and execute scripts for text manipulation. Evidence: In SKILL.md, it recommends using a Python script with the subprocess module to edit the map body in place. This involves the dynamic generation and execution of code based on patterns described in the skill.
[COMMAND_EXECUTION]: The skill relies on executing various shell commands via the GitHub CLI and Git to manage project state. Evidence: Procedural instructions include the use of gh issue reopen, gh issue view, gh issue edit, gh issue create, gh label create, and git commit/push across the SKILL.md file.

wayfinder-reversal