ClipCover

The problem

Short-form creators publish clips with an auto-selected frame as the cover, which is usually a blink or a blur. The existing tools generate an image rather than designing one, so the output looks like AI and not like a show.

What I built

• 01A pipeline that samples frames with ffmpeg, filters them through a blur and exposure funnel, detects faces, ranks the survivors with a vision model against a design playbook, and crops to 9:16.
• 02Audio transcription that returns a five-word hook, a highlight word, and an emotional tone, which then drives the layout selection.
• 03Subject isolation with background blur and rim lighting, composed through programmatic typography rather than a template.
• 04The commercial surface: accounts, Stripe subscriptions and credit packs, brand kits, saved layouts, batch mode, a Meta Pixel firing registration and purchase events, and English, Spanish, and Japanese localization.

Result

Launched publicly in July 2026 and used daily on my own channel. The design intelligence came from tearing down roughly 80 best-in-class podcast covers and sweeping 50 top creator accounts for palette patterns.