How I Built A Text To Commercial Generator Using Elevenlabs Gemini And Veo 2

This ElevenLabs blog post, first published May 29, 2025 and updated in September 2026, is a build log by Ryan Morrison of the ElevenLabs growth team. He describes a tool that turns a rough product idea into a short ad (the post gives both 20 and 30 seconds as the target) by passing work through several AI services in sequence, with a person able to edit or regenerate the output at every step.

The pipeline runs as an eight-step wizard. Google's Gemini 2.0 Flash first sharpens a vague idea into a specific product concept, then writes a script of four five-second scenes, each with voiceover lines, a video prompt and a sound-effect description. Each video prompt goes to Google's Veo 2 through FAL.ai's hosted API, since Morrison could not get Veo 2 working on his Google account. ElevenLabs' sound-effects API generates four mood variations per scene for the user to choose from, its text-to-speech API reads one combined voiceover, and FFmpeg mixes and joins everything. He reports about $10 of video generation for a 20-second clip without regenerations, and found background audio works best at roughly 30% volume under the voice. He says he wrote most of the code with Anthropic's Claude 4 Opus, over several days because of rate limits.

Morrison's main lessons are about holding a multi-service pipeline together rather than about any single model: prompt design took as much time as code, and he added fallbacks for each failure (stand-in scripts, placeholder videos, basic audio), local saving of expensive results, and resumable sessions because video generation is slow. The post publishes no source code and no measure of ad quality. It also predates Veo 3, as Morrison notes, so its video model and Gemini version are older than current options.