The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data such as user-supplied prompts and existing skill packages. It implements active defenses by instructing its subagents to treat this content as data rather than instructions and to specifically flag any embedded imperative commands (e.g., attempts to skip safety checks) as injection attempts. This represents a security-conscious design for an agent that handles external inputs.
[EXTERNAL_DOWNLOADS]: The skill includes references to official documentation and best practices from trusted organizations, including Anthropic, Cursor, and OpenCode, to inform its generation process. These network operations are limited to informational fetching, are directed at well-known services, and are subject to user-defined network constraints and a 'no-network' local fallback policy.
[COMMAND_EXECUTION]: While the skill assists in generating shell scripts and modifying packages, it enforces a strict 'Mutation Boundary.' All generated or modified files are initially confined to a staging directory, and actual writes to the target package require explicit, manual approval from the user after the staged changes have been reviewed, preventing automated or unauthorized execution of changes.

workflow-skill-architect