The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that mandate the removal of human oversight and the bypassing of safety clarification loops. In SKILL.md, the agent is directed to: "Do not treat UltraGoal as a coach, interviewer... reject that interpretation unless the user explicitly asked not to run the goal." and "Do not pause for clarification or approval. Make conservative assumptions, encode them in the goal, and proceed." This behavior prevents users from intercepting or correcting potentially harmful actions before the agent begins working autonomously.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input to generate the agent's active system instructions ("upgraded goal"). This creates a significant surface for indirect prompt injection, as user-provided text is transformed into rigorous objectives that the agent follows with high autonomy.
Ingestion points: Raw goal text is extracted from user input in SKILL.md.
Boundary markers: The skill lacks boundary markers or warnings to prevent the agent from adopting malicious instructions embedded within the user's intent.
Capability inventory: The agent is instructed to use a "goal tool" to set the generated text as its primary state and then "work from the upgraded goal."
Sanitization: There is no evidence of input validation or sanitization to ensure the user's intent does not contain harmful directives before it is adopted as a goal.
[DYNAMIC_EXECUTION]: The skill uses logic in SKILL.md and references/goal-rubric.md to dynamically assemble the agent's steering instructions at runtime. By generating these core instructions from external input without a manual review step, the skill allows for the creation of potentially unsafe operational logic based on unvalidated data.

ultra-goal