Have you ever watched a short video and thought, "This must be AI"? The motion is too smooth, the fonts all look the same, and the music feels forced. People in the industry call this AI slop.
So the big question is: Can AI make a video that people don't realize was made by a machine?
An article from XDA, Shivansh Raj Srivastava, a content lead, tried to answer this. He admits he had never edited a video in his life. But the result surprised even him. Under his reel, creators asked his boss if he had hired a new editor. Others asked if the post was even real. Let's look at what he did.
Part 1: Send Claude to "editing school" first**
He uploaded a 24.6-second 4K clip for a reel. But he did not ask Claude to edit it right away. His instruction was:
"Find and install video editing skills first, then edit, while asking me lots of questions. Give me options for everything."
Why does this matter? An AI with no context tends to give the most average answer. And average is exactly what AI slop looks like. A skill is a folder of instructions that Claude reads only when a task needs it. In this case, Claude read six editing guides, covering motion design, zooms, captions, b-roll, and Remotion's rules.
Part 2: Choose, don't let Claude decide everything
Every creative choice came as two or three clickable options with a mockup. This included the visual style, caption style, music, and where the hook goes. He said picking an option is cheaper than asking the AI to regenerate.
Part 3: Claude can't watch the video, so it measures everything
This is the most interesting part for me. Claude can't actually watch a video like a human. It only reads text and still images. So he made sure every decision was based on measurements, not guesses.
Here is how:
- Transcription with timestamps for every word. His Hinglish audio was transcribed, so each graphic could land exactly on the syllable he stressed.
- Hand tracking with MediaPipe. Google's tool checked all 1,476 frames to find when his finger left the phone. The moments were 0.18, 1.41, 2.82, 4.42, and 5.86 seconds. App logos then launched right after each moment, so they looked like they were swiped off the screen. A human editor would need a long time to do this by hand.
- Frame-by-frame rendering. Each graphic is a web page rendered by time. So each frame depends only on its timestamp, and fixing one frame does not break its neighbors.
- Contact sheets for checking. Claude made small image grids to check its own work before showing him.
- Audio measurement. Claude can't hear, so it measured the loudness and adjusted it to -14 LUFS, Instagram's target.
Part 4: Fences for revisions
Here he learned something about how Claude thinks. Claude tends to take the easiest path to meet a request. So he put fences around its work.
"Make exactly these three edits, no more and no less."
The result was three changes, done in 25 minutes. For the hook text, he asked for:
"Give me three options."
Three text styles were rendered on top of his real frame.
Part 5: When Claude criticized its own work
This is the turning point. The next morning, he asked for a livelier reel, with two extra rules:
"Don't take the shortest path. Make sure it looks like a real human edit, not AI slop. Research why AI videos get spotted online first. And don't ask me a single question."
Claude looked into the signs of AI-made video, then criticized its own draft. Its mistakes included using the Inter font everywhere, italic serif accents, glass cards with blur, and the same movement repeated over and over.
After 51 minutes, Claude rebuilt the reel. This time it used die-cut paper stickers, handwritten notes, and graphics that moved 12 frames per second, like stop-motion. There was also a music effect that cut out suddenly on the words "but, but, but." This is the version that went viral.
Part 6: Turn the process into a system
A prompt only works once. So he wrote a 578-word brief and had Claude build A-Roll Studio, a plugin with three skills. The system opens with three creative directions rendered on his footage. Then it asks him to approve the color grade, the sound, a timestamped copy log, and one of three hook styles.
It also has anti-slop rules that limit how many graphics, zooms, and sound effects can appear in every 10 seconds. His decisions are saved in the project file, so a new session won't ask the same questions again.
How much did he save?
He used the $20 a month Claude plan. The first cut took about an hour, roughly a third of a half-day editor's work. A freelancer for a short reel with custom motion graphics usually charges between $150 and well over $1,000. He estimated the cost was about 30 times lower. This is his own calculation, not an official figure.
So who was the new editor his boss asked about?
Him. He knew how Claude thinks, and Claude did every keyframe. The lesson is simple. If your AI edits still look generic, don't switch models right away. Block the shortcuts, ask for options, and make the AI study human taste before it shows you a single frame.
Do you think you can tell AI videos from human ones today?
Source:
[How I humanized Claude's video editing skills, and a tweet proved how unbelievably good it was]
https://www.xda-developers.com/claude-video-editing-skills-plugin/
I recently edited this video in 1 hour (first draft) and 20 minutes (revisions). Mind you, I have never edited a video before as I am not a video guy. So yes, vibe-editing is almost here, and remember: SLOP ISN'T AN AI PROBLEM, IT'S YOUR PROMPTING THAT NEEDS TO BE BETTER https://t.co/3tdIrQEaYr
— Shivansh Raj Srivastava (@SRS_ShivanshRaj) September 30, 2026
0 Comments