We are a fast-growing AI video production company building automated pipelines that transform reference videos into fully produced short-form content at scale. We're looking for someone who thinks deeply about how language models behave, knows how to engineer prompts that produce consistent, structured, high-quality output — and can wire that work into n8n automation workflows. The core of this role is prompt engineering. Everything else supports it. What you'll be working on: You'll own and improve the prompt layer of a production pipeline that takes a reference video through Gemini video analysis, AI scriptwriting, fal.ai Veo3.1 video generation, ElevenLabs TTS, and lipsync. The pipeline is built and running in n8n. Your primary job is to make the AI output better — more accurate, more consistent, and more aligned with what the video generator needs to produce great footage. Responsibilities: Write, test, and iterate on system prompts and user prompts for Gemini and other LLMs — with a focus on getting structured JSON output that passes strict quality validation on the first attempt Diagnose why the model produces bad output (re-describes scenes it shouldn't, ignores word budgets, invents details not present in the reference) and fix it at the prompt level Design prompt structures that give the model the right context, constraints, examples, and self-check instructions to produce reliable output without constant retrying Work inside n8n to connect prompt improvements to the validation and retry logic already in place Understand how AI video generators interpret scene descriptions and write prompts that produce motion, not frozen frames Identify when reference video content may trigger content moderation systems and reframe scene descriptions to accurately represent the concept without triggering flags Reduce retry rates and API costs by improving first-pass output quality through better prompting What we're really looking for: Prompt engineering is the primary skill. We need someone who has spent real time studying how LLMs behave — not just someone who has written a few system prompts. You should be able to look at a bad model output and immediately identify whether the failure is a context window issue, an instruction conflict, a missing example, a formatting constraint violation, or a model capability ceiling. And then fix it. n8n experience is helpful but secondary. If you are an exceptional prompt engineer who has not used n8n before, we are still very interested. The workflow is built. The nodes are connected. What we need is someone who can dramatically improve what the AI produces inside those nodes. You must have: Deep, hands-on prompt engineering experience with large language models — Gemini, GPT-4, Claude, or similar Proven ability to engineer prompts that produce structured JSON output reliably at scale Experience building prompts with system instructions, few-shot examples, chain-of-thought reasoning, and output validation hints Understanding of why models ignore instructions and how to fix it — through prompt restructuring, constraint layering, or example injection Ability to read JavaScript to understand what the validation logic expects, even if you are not writing it yourself Strong analytical thinking — you treat prompt failures as debugging problems, not guesswork Nice to have: Experience with n8n or similar automation tools (Make, Zapier, etc.) Familiarity with Gemini's multimodal capabilities — video or image analysis Experience with AI video generation APIs such as fal.ai, Runway, or Kling Understanding of how text-to-video models interpret scene descriptions and their content moderation behavior Familiarity with ElevenLabs or TTS workflows