Agent Opus
Agent Opus turns a prompt into a finished video. I redesigned the editing experience around one predictable framework, making it easier to refine AI-generated videos without relearning how editing works. Editor entry grew from 13% to 37%.
AI Video Agent
Prompt Editing
Company
Opus Clip
Role
Product Designer
Tools
Figma
Time
Dec 2025 - Jun 2026
CONTEXT
Generating the video was never the hard part.
Users could reach a complete draft in minutes. But “close” was rarely good enough, and the editor gave them too little confidence or control to finish. Generation had solved the blank page; refinement had become the real bottleneck.
USER RESEARCH
Users needed editing to feel safe.
I spoke with creators who generated a video but abandoned the edit.
Their complaints looked different, but pointed to the same three needs: understand what the AI had built, trust what would change, and control the details without starting over. Those needs shaped the two challenges that followed.
PRODUCT DECISION
Five engines didn’t need five editing systems.
Agent Opus generated content through five different engines, each with its own editable properties. I initially explored an editing model for each, but maintaining five systems would multiply both product complexity and user learning. I pushed for one shared framework, starting with GenAI.
CHALLENGE 01
Change one frame and two shots move.
GenAI broke the mental model of traditional editing. Shots were generated between keyframes, and neighboring shots could share the same frame. Changing that shared frame could reshape both shots, but the editor gave users no way to see the dependency before they acted.
EXPLORATION
Showing everything was as unusable as showing nothing.
Each direction traded transparency against complexity. I kept reducing how much of the AI workflow the editor exposed until only the dependency users needed to make the next decision remained.
Trust came from predictability, not transparency.
CHALLENGE 02
One wrong detail shouldn’t cost a whole regeneration.
Making the AI’s structure predictable solved only half the problem.
Small fixes, a mistimed caption, the wrong music, a missing CTA, still required regenerating the video, adding wait time and risking changes elsewhere. Users needed a way to fix the details without starting over.
EDITING MODEL
Some edits should regenerate. Most shouldn’t.
I split editing into two paths. Global edits re-prompt the model and return a new video; local edits apply instantly without touching anything else.
I scoped the first local layer to text, captions, background music, and CTA, changes users needed often, but none of which required the model to run again.
DEEP DIVE
Choosing a track and tuning it are two different jobs.
BGM put the model to the test. Choosing a track and refining it required very different levels of attention, but my first pass exposed both at once and felt like a settings panel. Separating them into two layers kept selection fast while preserving deeper control when users needed it.
REFLECTION
The model will change. The way people edit shouldn’t.
The model kept changing throughout this project: new structures, capabilities, and content types kept arriving.
I learned not to design around its current edges, but around a stable editing logic that new capabilities could join without asking users to relearn how editing works.
OUTCOME
Nearly three times as many people opened the editor.
More users entered the editor, and more stayed to finish. The stronger signal was behavioral: people started making a second edit instead of leaving after the first.
