The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses high-pressure language ('MANDATORY', 'CODE RED') and defines itself as a 'binding directive' to force the agent into a specific behavioral mode, attempting to override default operational guidelines.
[DYNAMIC_EXECUTION]: The instructions require the agent to use JavaScript eval as its primary execution environment for logic and automation, necessitating the generation and execution of code at runtime.
[COMMAND_EXECUTION]: The agent is directed to use shell-based tools including curl, tmux, and kill to manage system processes and perform network requests during quality assurance phases.
[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves ingesting and executing code, tests, and web content from untrusted external sources (ingestion points) without explicit boundary markers or sanitization, combined with powerful capabilities like eval and shell access.
[DATA_EXFILTRATION]: The requirement to 'ALWAYS ACTIVELY RECORD' every decision and outcome into a local temporary file creates a risk where sensitive information harvested during the coding process could be inappropriately persisted.

ultrawork