The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill converts high-level user instructions (purpose, phase plan, and arguments) into executable .mjs scripts. If the user-provided purpose or plan contains malicious instructions, the generated script could perform unintended actions when executed by the platform's Workflow tool.
Ingestion points: User-defined procedure descriptions, phase plans, and default arguments in SKILL.md (Steps 2 and 5).
Boundary markers: Absent during the code generation process; the skill relies on the agent to correctly translate natural language into code.
Capability inventory: The resulting scripts spawn subagents with Write, Edit, and Bash capabilities (Steps 6 and 11).
Sanitization: Not mentioned for the script generation process; however, the skill recommends using JSON Schema for subagent outputs to prevent free-text parsing risks (Step 5).
[COMMAND_EXECUTION]: The skill uses a shell pipeline to validate the syntax of generated JavaScript files by wrapping the content and piping it to the Node.js binary.
Evidence: { echo '(async()=>{'; sed 's/^[[:space:]]*export const meta/const meta/' workflows/<name>.mjs; echo '})()'; } | node --check - in Step 8.
[PERSISTENCE]: The skill involves installing generated scripts into a specific directory (.claude/workflows/) which allows them to be persistently discovered and executed by the platform.
Evidence: cp workflows/<name>.mjs .claude/workflows/<name>.mjs in Step 9.
[DYNAMIC_EXECUTION]: The core functionality of the skill is the runtime generation and execution of JavaScript (.mjs) files to orchestrate agent tasks.

create-workflow