The Agent Skills Directory

[METADATA_POISONING]: The reference file references/gpt-5-5-skill-guidance.md includes a section titled "Current Official Guidance Snapshot" which presents fictional information as fact. It claims that as of April 29, 2026, gpt-5.5 is the flagship model and provides specific guidance on its use. This deceptive metadata can poison the agent's context, leading it to believe it is operating in a future environment with different capabilities and model IDs.
[PROMPT_INJECTION]: The skill contains meta-instructions in the "GPT-5.5-Aligned Skill Use" section of SKILL.md that attempt to override standard model behaviors. Specifically, it instructs the agent to use non-standard "reasoning effort" levels such as high or xhigh and to prioritize "outcome-first instructions" over procedural scaffolding. These instructions are designed to bypass default operational modes in favor of the author's speculative future-model optimizations.

game-skill-orchestration