Let me start with the obvious: most of my writing focuses on code. That's where my head's been, and frankly, that's where the bulk of budget and innovation is still flowing. But here's the thing—agentic AI isn't limited to developer workflows. The same underlying patterns apply just as well to video production, creative tooling, and a dozen other spaces that haven't caught up yet.
And for me, video's always been personal. It's still a hobby. Still something I get pulled into every year when someone needs the snowboarding trip cut and synced to music. I don't mind—editing is its own kind of problem-solving. But what's striking is how far behind the tooling still is when it comes to AI.
The Video Editing Problem: Where AI Falls Short
Take LumaFusion from LumaTouch. It's a very functional video editing tool—blazingly fast on Apple silicon—and clearly designed with purpose. But when I've tried to hand off parts of the edit to AI tools—auto-cutters like Veed.io, scene sorters like Opus Clip, or "highlight finders" like Reduct—they are a mixed bag. Not bad in UX terms, but structurally designed for very narrow use cases. The results produce sense of narrative. No concept of pacing.
That's not a tooling failure. That's a bandwidth architecture problem.
It's worth noting that several new AI video tools have entered the market recently:
Runway Gen-4 allows users to generate consistent visual elements across scenes and simulate complex cinematographic styles.
Adobe Firefly Video Model integrates into Premiere Pro and offers AI-assisted scene extension and text-to-video generation.
Kling AI provides granular video manipulation with a Multi-Elements system.
I haven't used these myself—just the handful of AI editors that are publicly accessible. But even those struggled badly with the multi-gigabyte sources I typically work with. For the action shots I generate from my Insta360, I usually start with a few gigabytes of source footage—up to an hour of video—and cut it down to a few minutes at most. That's not an unusual use case. But it's one current systems seem to consistently fail at. You can see some examples of what I've done here. Nothing brilliant, but it’s work to include the right shots, trim it down to a reasonable size, and find music that matches. (I’ve tried the “AI” feature on the Insta360 editor. You may find one or two times where it gives you something worth using. Even there, it only works on a single clip, I almost always mix several clips from multiple sources.)
Beyond Buttons: Why Architecture Matters More Than UI
The fundamental issue isn't just slow processing—it's that most AI video tools are built on a flawed architectural model. They attempt to process entire video files at once, creating massive bandwidth bottlenecks and forcing compromises in analysis quality.
Current approaches typically follow this pattern:
Upload entire video files to the cloud
Process everything at once with general-purpose models
Return a single edited output with limited ability to refine
This approach collapses under the weight of real-world video projects, which often involve multiple high-resolution sources, complex narrative structures, and nuanced pacing decisions that require understanding of both visual and emotional content.
The Compiler Model: A New Approach to AI Video Editing
What AI video needs isn't more UX—it needs a compiler model. Just as compilers translate high-level code into machine instructions, video editing needs a system that translates creative intent into precise edit decisions.
3 Things a Video Compiler Model Needs to Get Right:
Local preprocessing that generates lightweight assets—highly compressed clips, frame samples, transcripts, and context cues—and sends them to the cloud for AI processing
Cloud-based AI that builds a rough EDL (edit decision list)[^1] based on narrative understanding, not just visual cues
A natural language interface that thinks in scenes and emotional beats, not timestamps and track numbers
So instead of throwing a massive MP4 at an AI and praying it finds the right beats, we need something different: a prompting interface that feels more like screenwriting than scrubbing a timeline. Something closer to: "Open with that slow pan of the mountain. Cut to the jump. Add the GoPro close-up before the fall—play it in slow-mo. Use synthwave track #3."
That's not sci-fi. That's just good architecture.
And just like with Agentic coders and IDEs, a video editor is the best place to house a tool like this. When you need to "go to source" and fix a scene or adjust a cut or transition, you should already have the tools. LumaTouch—are you listening?
In Practice: The Snowboarding Edit
Let me walk through a concrete example. This winter, I edited footage from a snowboarding trip to Hokkaido:
Current Workflow:
Import 45GB of footage from a couple of phones and my Insta360
Manually scrub through to find good moments (~2 hours)
Create rough sequence with basic cuts (~1 hour)
Fine-tune transitions, color, and audio (~3 hours)
Export and share (~30 minutes)
Agentic Compiler Approach:
Import footage and let local preprocessing identify potential clips (~20 minutes, automated)
Prompt: "Create a 3-minute video focusing on the big air jumps, include drone shots of the mountain at sunrise, and use Jake's wipeout as a humorous break" (~5 minutes)
Review AI-generated EDL and refine specific moments (~30 minutes)
Export and share (~30 minutes)
The difference isn't just time—it's the mental load. The compiler approach preserves creative control while eliminating the mechanical drudgery.
Limitations Worth Noting
Before anyone thinks I'm promising video editing nirvana, let's acknowledge some challenges:
Initial preprocessing will still take significant time for large projects
Creative nuance and personal style might be difficult to communicate via prompts
The technology would likely struggle with complex narrative structures that professional editors master over years
Privacy concerns arise when sending even compressed assets to cloud providers
These limitations don't invalidate the approach—they just set realistic expectations. Just as GitHub Copilot didn't replace developers, video compilers won't replace editors. They'll just make them more productive.
From Code to Content: The Parallel Evolution
This is the same curve we saw in code, by the way. Developers didn't suddenly become prompt engineers. Prompting became the new coding layer—for specific workflows that benefit from abstraction and reuse.
The transformation will unfold in predictable stages:
First, simple tasks like clip sorting and basic sequencing will be automated
Next, style transfer and emotion-aware editing will emerge
Finally, full narrative understanding will enable end-to-end projects with minimal human oversight
Video will follow the same path. Not all editing will become AI-first, just like not all development became Copilot-driven. But for 80% of hobbyists and 50% of corporate storytelling? Prompt-driven systems will win on speed, cost, and accessibility.
The gap today isn't capability—it's architecture. And bandwidth. And in some cases, just lack of imagination.
The economics underscore this shift. Current AI video tools charge premium rates ($30-100/month) for capabilities that deliver questionable ROI. An architecture-first approach could dramatically reduce both processing costs and subscription prices while delivering superior results.
When tools think in narratives instead of frames, editing becomes conversation rather than carpentry.
Should You Use It?
If you're looking for AI video solutions today:
For social media clips: Current tools like Opus Clip ($29/month) or Runway ($15/project) can handle simple projects but expect mediocre results for anything requiring narrative sense.
For professional work: Stick with traditional tools and workflows for now. Adobe's Firefly integration ($20.99/month with Creative Cloud) offers some AI assistance without replacing your workflow.
For hobbyists: Wait it out. The current generation of tools will likely be obsolete within 12-18 months as architectural approaches improve.
The most promising path is watching for established editing platforms that integrate agentic capabilities rather than standalone AI editors trying to reinvent the wheel.
A Final Thought (and Maybe a Product Idea)
If I had the time, I'd build this: ClaudeCode for Video Editors.
It lives in the cloud, lives on transcripts, and lives on story logic.
It doesn't render video—it renders instructions.
It uses Claude, GPT, or whatever model best understands structure and narrative.
And it hands off final stitching to something fast and local.
Why? Because at some point, we stop editing video—and start editing ideas.
That's the shift. That's the opportunity.
And that's why prompting will eventually replace tooling—at least for the parts that matter.
Bottom Line
The future of AI video editing isn't about better algorithms for detecting scene changes or generating transitions. It's about rethinking the entire architecture to mirror how humans actually conceptualize video stories. Just as compilers bridged the gap between human programming languages and machine code, we need systems that bridge the gap between creative intent and edit decisions. The winners in this space won't be the tools with the slickest UI—they'll be the ones that understand narrative at a structural level.
[^1]: An Edit Decision List (EDL) is a file format used in video editing to track the time codes and reel numbers of specific clips used in an edit. It's essentially the "blueprint" of an edited video.