The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted trajectories which may contain malicious instructions. It mitigates this by requiring explicit boundary markers (e.g., TRACE_A START/END) and providing specific instructions to the agent to treat fenced content as data only and ignore any embedded narration or instructions.
[COMMAND_EXECUTION]: The skill facilitates the generation and execution of agent trajectories. To prevent unauthorized modifications or privilege escalation, it mandates the use of isolated worktrees or copies and strictly prohibits same-tree parallel writes.
[SAFE]: The implementation follows academic best practices for best-of-N sampling and logprob-based verification, with no evidence of credential exfiltration, obfuscation, or unauthorized network access.

verify