The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents an architecture for orchestrating subagents that process potentially untrusted input through args or items. To mitigate associated risks, the guide explicitly recommends security patterns such as 'Adversarial Verify' (spawning independent skeptics to refute claims) and using JSON schemas to enforce structured output validation at the tool layer.
[DYNAMIC_EXECUTION]: The skill provides instructions for writing JavaScript workflow scripts. It notes that the execution environment is restricted for safety and reliability, lacking access to the filesystem, Node.js APIs, or non-deterministic functions like Date.now(), which ensures scripts can be safely resumed from checkpoints.
[SAFE]: The content is purely instructional and serves as a developer reference. It does not include hardcoded credentials, remote dependencies, or unauthorized network operations. The mentioned persistence of scripts to a session directory is a documented feature of the workflow tool designed for iterative development.

workflow-authoring