The Agent Skills Directory

[SAFE]: The skill implements a secure execution model by teaching the agent to author orchestration scripts that run in a restricted environment. A provided linter (scripts/validate-workflow.mjs) automatically detects and blocks the use of sensitive Node.js host APIs such as fs, process, and require within the orchestrator script.\n- [SAFE]: External references, including the installation repository github.com/ray-amjad/claude-code-workflow-creator.git, are legitimate vendor resources used for providing the skill's reference materials and examples.\n- [SAFE]: Instructions emphasize safety best practices, such as implementing budget guards in loops and ensuring determinism to support reliable execution and resume functionality.\n- [PROMPT_INJECTION]: Workflows authored using this skill may ingest untrusted external data (e.g., CSV files or logs), representing a surface for indirect prompt injection where malicious content in those files could influence subagent behavior.\n
Ingestion points: Data enters via the args global variable or through file reads within subagent prompt templates (e.g., in the customer-feedback-theme-extractor.js example).\n
Boundary markers: Code templates do not consistently mandate the use of XML tags or explicit 'ignore' instructions when interpolating external data into prompts.\n
Capability inventory: Subagents managed by the workflows have access to standard agentic tools, including shell access and filesystem modifications.\n
Sanitization: The skill provides patterns for normalizing data structure (e.g., JSON parsing) but does not enforce content-level sanitization for ingested data.

workflow-creator