The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructs the agent to treat content inside EVO DIRECTIVE banners as "user-authoritative" and to "Follow it verbatim". It further includes a self-referential claim that "This is not tool-output prompt injection", which is a pattern used to encourage agents to ignore their internal safety filters when processing external data.
[COMMAND_EXECUTION]: The protocol utilizes the evo bash command to execute arbitrary shell commands within a remote container environment (--backend remote), as well as the evo run command to execute benchmarks.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon external data including benchmark logs, task traces, and user-supplied directives spliced into the conversation context.
Ingestion points: EVO DIRECTIVE banners in the conversation, task traces accessed via evo traces, and experiment outcomes from outcome.json.
Boundary markers: Content is wrapped in [EVO DIRECTIVE] and [END EVO DIRECTIVE] banners.
Capability inventory: The agent has the ability to execute shell commands (evo bash), run scripts (evo run), modify files, and write annotations.
Sanitization: No sanitization is specified; the agent is explicitly instructed to follow the directive content "verbatim".
[DYNAMIC_EXECUTION]: The skill directs the agent to dynamically discover and load additional functional scripts ("category skills") at runtime based on the output of CLI commands like evo config get task-skills.

subagent