Testing an AI Video Feature Before Engineering Built It
A working prototype anyone can open, and a spec drawn from it.
Problem
Product-level video is slow and manual for e-commerce and marketing teams. It is hard to scale across catalog ads.
Optifeed already holds the product data, images and channels. So AI video was a natural next step.
The risk was not whether AI could make a video. It was whether users would trust the flow enough to review, approve and export it themselves.
So the flow had to keep users in control before anything reached an ad channel.
What I did
- Owned the product direction: discovery, definition, and the first version’s scope.
- Built the working prototype with Claude Code and Codex. How I use AI →
- Ran two internal stakeholder feedback sessions, then iterated.
- Handed engineering a solution spec drawn from the prototype.
The flow
- Start. Pick products and set up the campaign.
- Configure. Choose a template or scenario, including the Textile/Fashion branch.
- Commit. Confirm before spending tokens or waiting for output.
- Control. Review, approve, reject or regenerate, then choose where to export.
What feedback changed
Approval is mandatory.
No video moves toward a channel on its own.
Cost before commitment.
Token cost and wait time show before generation starts.
One category first.
Textile/Fashion became the first engineering scope.
The user picks the channels.
Export needs an explicit choice.
Where it stands
Engineering received the live prototype and a solution spec: product logic, key decisions, visible states, and what was out of scope.
The feature is meant for Optifeed’s existing B2B customers.
When my engagement ended, the first product build was still in progress. I did not see it launch, so there is no usage, adoption or revenue data.
