AI engineers will spend 5, 10, sometimes 20 hours building one n8n workflow. In this video, I tested whether we can actually automate that build process with AI by having three different tools generate the same n8n workflow from scratch—and then scoring what they produced. I ran the exact same prompt through Claude (Sonnet 4.5 with an n8n MCP connector), “anti-gravity” using Gemini 3 Pro, and n8n’s native AI workflow builder. For each one, I timed it, saved the JSON output, and judged it on correctness, error handling, documentation, and (when AI nodes were involved) prompt quality. I also paid attention to whether the tool asked clarifying questions, because vague prompts are the norm in the real world. What I found was pretty straightforward: speed doesn’t matter if the workflow doesn’t run. Claude asked strong clarifying questions and produced a nice README, but the actual workflow wiring/logic was off enough that I didn’t expect it to execute correctly (and it didn’t match some requirements like Gmail). Anti-gravity was fast, but had weak documentation, minimal error handling, and even basic trigger settings were wrong (like running hourly instead of daily). n8n’s native builder asked the best questions and produced the best AI prompt and structure overall, with real error-handling branches—but it still had correctness issues (like not properly connecting all RSS sources). The takeaway: AI can get you a big head start, but you still need an engineer’s eye to validate wiring, triggers, credentials, and edge cases before anything goes near production.