AI Doesn't Replace Expertise. It Makes Expertise Portable.

This weekend I was in a hotel room in Los Angeles. No ring light, no studio mic. Just my iPhone propped up on top of my laptop wedged against the desk.

I recorded a three minute talking head video.

If you know me, I never do anything with videos except publish them. No post-production. No color correct. No tweaking audio. Barely any captions. And never the work to do that zoom thing that YouTube creators use today.

I still didn't do any of that.

But this time the video came out different. I typed a single command, and went down to grab lunch. When I came back - silence was trimmed, zoom sections were placed based on my words. Colors corrected. Audio mastered. Captions added. And the video resized from 4k to HD.

No Final Cut. No Premiere. No Capcut or Descript.

It's a GitHub repo with eight markdown files. Eight Claude Code skills that automate video post-production from start to finish.

No application code. No interface. No dependencies beyond ffmpeg and a local whisper model. Just structured instructions that Claude Code reads and executes.

Here's the demo video if you want to see what comes out the other end. Remember — iPhone. Hotel room. Laptop tripod.

What most people think AI video editing means

When people hear "AI video editing," they picture one of two things.

The first is the SaaS approach. Upload your video to some platform, click some buttons, wait for their servers to process it, download the result. Maybe you get a free tier with a watermark. Maybe you pay $30 a month. The platform decides what "good" looks like.

The second is the code approach. Write a Python script. Wire up ffmpeg commands. Debug your way through audio stream index mismatches and filter graph syntax until something works. Then do it all over again the next time you have a different video.

Both approaches share the same assumption: that the intelligence has to live somewhere specific. Either in a company's servers or in your code.

I tried the opposite.

What if the intelligence lived in the instructions?

The problem was never "how do you remove silence from a video." FFmpeg has done that for years. The problem was orchestrating all the steps together, making editorial judgments about which moments deserve emphasis, and doing it consistently without babysitting each step.

So instead of building an app, I wrote Claude Code skills for video editing. Each markdown file describes one step. Not vaguely — specifically. With parameters, with error handling, with the exact ffmpeg filter chains that produce good results for talking-head video.

Step 1 detects silence longer than half a second and cuts it down to 0.3 seconds of natural pause. Step 2 transcribes the audio with whisper and breaks the transcript into timed sections. Step 3 is where it gets interesting.

The step where AI actually earns its keep

Step 3 reads the transcript and makes editorial decisions.

It looks at each section of what you said and categorizes it as "normal," "emphasis," or "critical." Normal content — your setup, your transitions, your connective tissue — stays at full frame. Emphasis moments — your key supporting points, your rhetorical questions, the places where your energy rises — get a 1.25x zoom. Critical moments — your thesis statements, your punchlines, the parts you'd clip for social media — get a 1.6x zoom.

The instructions tell Claude Code to aim for roughly 40% normal, 35% emphasis, and 25% critical. No section shorter than 3 seconds (that looks manic) or longer than 7 (that looks static).

This is the part that used to require a human editor scrubbing through footage making judgment calls. Now those judgment calls are encoded as instructions — what makes something "critical" vs. "emphasis" vs. "normal" — and Claude Code applies them to the transcript.

Then step 4 takes those labels, detects where my face is using OpenCV, and renders each section at the appropriate zoom level, cropped and centered on my face. Step 5 corrects the colors with a warm-punch preset tuned for indoor talking-head video. Step 6 runs a full audio mastering chain — highpass filter, presence EQ boost, compression, loudness normalization. Step 7 burns in captions with a specific font, size, and positioning.

One command. All seven steps. Time to grab a Diet Coke and celebrate.

Why markdown files and not an app

I could have built an app. A SaaS. A CLI tool. But I did it this way for three reasons.

First, I wanted to be able to let friends who might use this change things. They weren't going to be coding stuff. So giving them text files means they can tweak the different choices I made really easily.

Second, I had already created a skill graph for Your Content Agent and wanted to try another one. Eight files instead of sixteen. Video creation instead of article creation.

And third, I'm becoming a big fan of showing people how to turn their expertise into something they could sell. And this is a great example.

What this means if you're paying attention

I'm not a video production expert. I'm the guy who publishes raw video without touching it. But I've been going deep on Claude Code skills — building them, testing them, figuring out what they're good at and where they fall apart. And what I keep finding is that you can encode expertise you've picked up into a set of instructions and get remarkably good results.

The people who built their careers on specialized execution — the video editors, the audio engineers, the data pipeline builders, the QA specialists — are watching AI tools get close enough to "good enough" to be concerning.

But here's what I think they're missing: the most valuable thing they have isn't the ability to run ffmpeg commands. It's the knowledge of which commands to run, in what order, with what parameters, and when to break the rules.

When you encode expertise as Claude Code skills, it becomes something genuinely new. Not a tool. Not an app. Not a SaaS product with a monthly fee. A portable set of documents that makes any AI agent capable of doing the work the way an expert would do it.

The specialists who figure this out aren't getting replaced. They're becoming the people who teach the machines. And the teaching artifact is a markdown file.

Want to see what that looks like in practice? The repo is public: github.com/chrislema/videoeditor. And the demo video shows exactly what comes out of an iPhone recording in a hotel room with no production setup.

If you want to see the same pattern applied to content creation instead of video production, check out YourContentAgent.com.