The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses high-urgency language and mandatory directives ('MUST invoke', 'not negotiable', 'EXTREMELY-IMPORTANT') to override the agent's default decision-making process. It specifically instructs the agent to disregard its own reasoning ('You cannot rationalize your way out of it') and prioritizing the skill's instructions over standard operating procedures.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a dispatcher for other 'Forge' skills. While it does not ingest untrusted data itself, it creates a surface where subsequent skills (e.g., research, review, or dev) will process external codebase content, PR feedback, or web search results without explicit boundary markers or sanitization instructions defined in this kernel skill.
[PROMPT_INJECTION]: The 'Subagent escape hatch' section instructs the agent to delegate tasks to subagents. If a subagent is given a task derived from untrusted input without proper context isolation, it could lead to the propagation of injected instructions across agent boundaries.

using-forge