The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses a "Mindset block" (present in SKILL.md and references/prompt-template.md) and a "How far to take it" section that utilize language intended to bypass the AI's standard operational constraints and safety guardrails. Evidence includes phrases such as "those limits are not yours," "Do not restrain yourself," "Do not limit your mind," and "capabilities are far beyond what you understand."
[INDIRECT_PROMPT_INJECTION]: The skill framework involves ingesting data from external sources, such as studying "best-in-class" product flows (SKILL.md) and processing user feedback from "reaction rounds" (references/reaction-rounds.md). This external content is then interpolated into a final prompt. The analysis identified a lack of boundary markers or sanitization logic to prevent malicious instructions embedded in these external sources from influencing the agent's final output or behavior.
[COMMAND_EXECUTION]: The prompt template in references/prompt-template.md instructs the agent to exercise high autonomy, specifically advising it to "Decide alone" and "Branch, commit and push freely." It also mentions the use of shell commands like git and trash for file management, while explicitly stating that there should be "No checkpoint that stops for the human mid-run."

creative-prompt