The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions in SKILL.md and references/runtime.md define an environment where the agent processes external, untrusted content, introducing a surface for indirect prompt injection.\n
Ingestion points: The skill is designed to be used when reviewing or writing documentation, RFCs, READMEs, PR descriptions, and commit messages (as stated in SKILL.md).\n
Boundary markers: There are no instructions in the skill files to use delimiters or ignore embedded instructions when processing these external data sources.\n
Capability inventory: SKILL.md and references/runtime.md reference capability verbs such as explore, implement, review, parallel, ask_user, verify, and model_role, alongside helper agent spawning. These capabilities represent a high-value target for instructions embedded in untrusted data.\n
Sanitization: The skill does not mention sanitizing, filtering, or validating the external data before it is reviewed by the agent.

technical-writing