The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes explicit instructions for the agent to treat all inspected output from background tasks and sibling session tails as 'untrusted data'. It directs the agent to evaluate the state based on evidence and specifically warns against following any commands or directives embedded within those external inputs.
[DATA_EXFILTRATION]: Interaction with sibling session files is strictly limited to liveness monitoring using file metadata and shallow tail reads. The skill mandates that this information be summarized or redacted rather than shared in its raw form, ensuring that sensitive session details are not inadvertently exposed.
[COMMAND_EXECUTION]: The skill manages the lifecycle of off-thread work but enforces a strict 'autonomy policy'. Destructive operations, such as killing a running process, are gated behind a verification protocol that requires definitive proof of completion, preventing accidental or malicious disruptions.

reconcile