Reference diagram

AutoCreative — AI Marketing Autopilot

Inspiration

Small businesses lose campaigns to time and budget — agencies are slow and expensive, and most AI tools only generate one piece of the puzzle (just copy, or just an image). We wanted an agent that takes a single brief and autonomously delivers a complete bilingual marketing package: strategy, platform copy with hashtags, a promo video with real narration, and a review step — with zero manual generation work.

What it does

A client fills a short form. In under 2 minutes, AutoCreative:

Generates strategy, hook, and bilingual copy with hashtags per platform using Qwen-Plus
Animates the product image into a promo video with native AI narration using wan2.6-i2v
Sends a review email with Approve / Request Changes buttons
On approval, generates the final video. On revision, Qwen rewrites the package and a new video is generated asynchronously.

How we built it

Google Apps Script as the orchestrator and Web App
Qwen-Plus (DashScope International) for all copy and the visual prompt — including an explicit audio direction so the video model narrates in the client's language
wan2.6-i2v for image-to-video with audio: true, generating real spoken narration directly from the prompt
Google Drive for asset storage and public image hosting
A human-in-the-loop approval flow built with time-based triggers to avoid Apps Script's 6-minute execution limit during revisions

Challenges we faced

Model availability and behavior varied wildly across DashScope regions — the same model name returned 404, 403, or worked perfectly depending on the endpoint (international, US, Singapore, China)
Getting real voice narration (not just background music) required treating it as a prompt-engineering problem: explicitly instructing the model what to say and in which language
Alibaba's content filter rejected innocuous prompts containing words like "sweat" or "struggle," requiring a sanitization layer
Google's corporate Workspace URLs broke our human-in-the-loop forms in subtle ways, costing hours of debugging parameter loss in GET requests

What we learned

That a working pipeline with known limitations beats a perfect design on paper — and that the fastest way to know what an AI model can actually do is to test it directly, not trust the docs.