Context
ReelUse takes one long video and returns five to ten short clips, strongest first, usually within 10 to 15 minutes. Each clip starts at the beginning of a thought and ends where the point lands, with captions and a title written for it. Customers paste a link in the browser or open the original file in the free Mac and Windows app, which keeps the full recording on their own computer.
Its customers are podcasters, churches and media desks: people with hours of footage and no evening to spare. My job is to make sure everything around the AI, from the release and the price to the help page and the email, earns the same trust as the clips do.
01
From signals to scoped work
I own the roadmap and prioritisation across web, desktop and mobile. Inputs come from three places: customer feedback, support conversations and usage data. My job is to turn them into features that are scoped tightly enough to ship.
02
Ship it to ourselves first
Every new feature ships behind a per-account beta flag. We switch it on for internal accounts and run real workloads through it, meaning real recordings and real lengths, not a demo file. Only then does it go to all customers.
Launch day isn't the day the code goes live. It's the day the help content and support answers are ready too. A feature that works but can't be explained isn't finished.
03
People judge quality, not the model
An AI that picks clips will always have an opinion about its own picks. I don't use that opinion as the measure. We judge quality with blind human reviews: real clips cut from the same recordings, rated Good or Weak by people who don't know which model picked them.
Test material is our own recordings plus one customer who opted in. Customer content is never used without consent, and that applies to testing too.
Live, we watch downloads and the feedback reports customers send. The next signal I've planned is the one that matters most: whether each clip actually gets posted.
04
The badge we pulled the same evening
We shipped a Top Pick badge to mark the strongest clip in each video, then measured what it was doing. Its score mostly ranked clips against each other rather than measuring quality, and over a quarter of videos got no Top Pick for reasons that had nothing to do with quality.
A badge that says "this one is best" has to be true. This one wasn't reliably true, so we withdrew it that same evening. I wrote about it.
05
Choosing the model
We compared the AI that chooses the clips across three providers, on quality, speed and reliability. In blind ratings the best models had roughly two in three picks rated good. The most capable and most expensive model scored worst.
The finding that shaped everything after it: run the same recording twice and only about half the clips repeat. So one comparison is noise. Model decisions need repeated human review, not a single test.
In October the live model slowed to over half an hour per video and failed often under load. We replaced it with a sibling model that chose clips in about two and a half minutes, failed no runs and rated best in our quality reviews. Separately, we redesigned the selection steps to cut AI calls per video by about 70% with no measured loss in quality.
06
Designing for the ways it breaks
Most failures weren't the AI being wrong. They were everything around it. Each one got a product change:
Video hosts refusing or slowing downloads was the main cause of failures under load. We now fetch each video once and spread downloads across more connections.
The AI provider running out of capacity. We switched models and planned an automatic backup.
Framing the wrong person, like a face in a slide photo or the slide screen itself. We moved to AI framing that identifies who is actually speaking.
Languages we can't transcribe yet. We only claim the languages we've proven.
In a load test of 40 long videos at once, 39 finished in about 35 minutes, producing over 300 clips.
07
The AI suggests, the customer decides
I shaped the core clip experience around one idea: the AI suggests, the customer decides. That meant selectable clip lengths, widescreen long-form clips for people who don't post vertically, and in-app trim and extend: the customer drags a clip's start and end along its transcript, even past the AI's original cut, without re-running the job. It went to every customer in October, after running on internal accounts first.
For teams and brands: saved brand presets with free logo and handle placement, and per-seat team branding so every editor's exports look like the same organisation.
08
Pricing that doesn't punish people for our mistakes
I defined pricing and packaging across three paid tiers and team plans, including the upgrade path and free-plan limits.
The rule I'd defend in any room: a failed job never counts against a customer's allowance. AI processing fails in ways the customer can't predict or prevent. If we charge for that, a technical problem becomes a trust problem. I wrote about why.
Underneath the tiers is a number: we measure the full cost of processing every video. That measurement informed plan design and the refund policy. Cutting AI calls by 70% and choosing a faster model were cost decisions as much as quality ones.
09
The help centre is part of the product
I led an audit of the help centre that found about 45 inaccurate claims and 40 undocumented features. An inaccurate claim sets the wrong expectation, and an undocumented feature is value nobody finds.
Fixing the pages was the easy half. The lasting change was a standard: every customer-facing release updates help pages, feature pages and support answers in the same change. More on that here.
10
One vocabulary, app to inbox
I designed the lifecycle email program: onboarding, activation, delivery, billing and re-engagement. It's grounded in research into how global consumer brands write to their customers.
Underneath it is one consistent product vocabulary across the app, the help centre and email. One thing, one name, everywhere.
11
Customer data is opt-in
I set the customer data policy: explicit opt-in before any customer content is used to improve the product, with clear retention rules for paid and free accounts. People upload sermons, interviews and family moments. The default has to protect them.
12
Release communication and research
Every release gets an in-app What's New entry and a changelog line, and every major feature gets its own page.
Behind the decisions is a research practice I built: competitor and market studies, AI model comparisons and user feedback, kept in a shared library so each decision builds on what we've already tested instead of starting from scratch.
Role details and figures from Ibukun's own account of her work; product details from reeluse.com.