Case Study · AI Short-Form Video Studio

1minutereel.com

An AI short-form video studio, designed and built solo with Claude Code in about a month — development started late Jan 2026, launched Feb 2026, refining since. Live and growing.

The Problem

I had just shipped 1minutefunnel.com — an AI-driven quiz funnel builder. The funnel half worked well. But every user kept running into the same wall: they needed a short video ad to run traffic into the funnel, and the existing tools for making one were terrible.

The category was full of products that wanted you to be a video editor first and a marketer second — timelines, layers, keyframes, render queues. A funnel builder asking a small-business owner to learn premiere-pro-by-another-name on the same afternoon they bought my product was not going to work. So I built the tool I wished my funnel users had.

It was also a two-year backlog item in my own head. I had wanted to build something like this in 2023 and 2024. I could not have, then. The models hadn't caught up, and the cost of any single render was too high to make the math work for a solo builder. By late January 2026, both had shifted enough that I could ship the whole thing in a month.

Constraints

• One person. No team, no engineering org, no design system to inherit. Claude Code as the entire build pipeline.
• Render compute is expensive. FFmpeg-orchestrated video generation costs real money per output. The UI had to respect that without making the user feel rationed.
• The user is not a video editor. Default to AI choices; keep tweaks shallow and obvious.
• Speed-to-first-output matters more than perfection. The product name is a promise: a minute.
• Mobile-first capture and viewing. Vertical 9:16 first, with horizontal and square as matrix-multiplied outputs.

The Approach

Three design calls drove the build. Each was constraint-shaped: by the workflow the product had to replace, by the user it had to serve, by the economics of generating video at scale.

Decision 1 — Integration is subtraction.

The product replaces nine specialized tools — ChatGPT, BIGVU, CapCut, Descript, Loom, Buffer, Testimonial.to, Bonjoro, Wistia — with one workflow. Each of those tools has hundreds of features I deliberately didn't replicate. CapCut's effect library. Buffer's full integration matrix. Descript's audio mixing. Every feature I cut was a feature someone, somewhere, will miss. Every feature I kept earned its place on the workflow's critical path.

The thesis comes from Larry Keeley: "In the 21st century, for the first time in the history of our species, innovation is less about the primary invention of the new and more about the elegant integration of the known."

The hard part of integration is subtraction. Nine tools, nine context switches, three hours per reel — collapsed into one tool, minutes per reel.

Proof of resonance. We put this exact framing in an ad — "9 tools. 9 context switches. 3 hours. For ONE reel." — and 83% of viewers watched to the end. Because it's not a pitch. It's their life. The workflow failed people, not the people.

Decision 2 — AI-curated content inside a locked surface.

Most video tools assume you are a video editor. The user of 1minutereel is not. They have a business to run, a message to share, a customer to talk to — and an exhausted hour somewhere between dinner and the dishes to fit recording into.

The design call was to remove the choices that don't matter, and multiply the time on the ones that do. The teleprompter auto-scrolls. Auto-trim handles dead air, filler words, false starts, and repeated takes. Auto-captioning works across thirty languages. CTAs auto-place inside safe zones across aspect ratios. The user supplies the content; the system curates the polish.

Same thesis as 1minutefunnel's locked layout: restraint as user respect. The fewer choices the user has to make about form, the more attention they can spend on substance.

Decision 3 — The transcript is the editing surface.

In most video tools, AI is an assistant — it suggests, the human decides. In 1minutereel, AI is the editor — it cuts first, the human reviews. Auto-trim runs across four categories on every clip: leading and trailing dead air, mid-clip pauses, filler words ("um," "uh," "like"), and repeated false starts.

The transcript becomes the editing surface. Delete words in the text, and the underlying video deletes with them. The user thinks in language, not timelines. That single shift — from scrubbing to reading — is what makes editing legible to a non-editor.

The trade-off is real: AI can't catch every edit a human editor would, and sometimes it cuts a beat that mattered. But for a non-editor user shipping fifty reels a month, the time AI saves is the product.

Key Screens

What Was Hard

Generating many videos at once meant orchestrating FFmpeg to keep render costs sane — and the order of operations was the nuisance. Captions, hook overlays, and CTAs each compose against the video at different layers. Stack them in the wrong order and the output is wrong, expensive, or both. It took several iterations to land the right composition pipeline.

Adjacent to that: the timing formula for engagement. After enough A/B passes, the answer landed at 3 seconds of hook + 3 seconds of CTA at the head and tail of every clip, both reinforced by captions on screen. That single timing decision — three plus three — moved completion rates more than any UI change I made. Design isn't always in the pixels. Sometimes it's in the seconds.

What I Learned About Designing With AI

The interesting design surface in an AI product is not the AI — it is the moment before the AI runs (what you let the user say) and the moment after (what you let them change).

The before is the prompt, the brief, the topic, the hook style — what the user provides as input. The after is the editable surface — what the user can change once the AI delivers. In 1minutereel the before is small and the after is curated: a topic and a hook style on the way in; transcript editing and simple text fields on the way out. The middle — the AI generation, the FFmpeg orchestration, the matrix multiplication into many cuts — is hidden by design. Users don't want to see the assembly line. They want the reel.

What I'd Do Differently

• Refine the core before adding capability. I added features faster than I refined the ones already there — and ended up with surface area I couldn't fully polish. Next time: lock the critical-path workflow at v1, run it through twenty real users, sharpen it to a knife's edge, then expand. The temptation to add the next obvious feature is real. It's almost always wrong.
• Build the meta-proof video first. The video walkthrough recorded in 1minutereel about 1minutereel should have been the launch artifact, not a post-launch one. The product that demos itself doesn't need a separate marketing motion.