The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides guidelines for building subagents that process untrusted data from user requests and external APIs. This creates an indirect prompt injection surface. The documentation explicitly addresses this risk by recommending the use of XML boundary markers to delimit instructions and providing specific prompts to treat external input as data rather than instructions.
[COMMAND_EXECUTION]: The documentation discusses the configuration and use of tools such as Bash, which allow agents to execute local commands. It incorporates security best practices by advising the principle of least privilege (deny-all by default) and restricting tool access based on the trust level of the data being processed.
[SAFE]: A thorough review of the skill and its references found no evidence of malicious patterns, obfuscation, unauthorized data exfiltration, or privilege escalation. The content is purely instructional and follows Anthropic's recommended safety guidelines.

create-subagents