Big Idea

The argument in one line.

A Claude Code editing skill built like a job description for a new hire, complete with Remotion templates and a feedback knowledge folder, can produce publication-ready short-form edits at batch scale that generic AI prompts cannot.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You regularly produce short-form content and want to cut post-production time without sacrificing visual quality.
You already use Descript for rough trimming and want to know what the next automation layer looks like.
You are comfortable running Claude Code locally and want a pre-built SKILL.md you can drop in and train over time.
You have tried generic AI editing prompts and found the output too inconsistent to publish.

SKIP IF…

You want a no-code button. This workflow requires Claude Code and an ongoing commitment to updating the skill file.
Your primary format is long-form narrative essay. The skill is optimized for short clips under 90 seconds.

TL;DR

The full version, fast.

The workflow splits into two phases: Descript strips dead air and bad takes then exports each clip as a raw scene file; Claude Code takes those files and runs a SKILL.md-driven pipeline that transcribes, fetches real images and logos, selects from a Remotion animation template library, tracks head position for dynamic zoom, and renders a low-quality preview for feedback before committing to full quality. Because every feedback round updates the SKILL.md knowledge folder, edit quality compounds session over session. The result is 18 shorts edited in a single batch, with the human only reviewing a preview and clicking approve.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:43

01 · Cold open: proof before promise

Shows the AI-edited short, states 10 hours of skill investment, frames the video as a how-it-was-done walkthrough.

00:43 – 01:44

02 · Descript pre-processing

Word-gap shortening, retake removal, scene-by-scene export produces 18 raw MP4s ready for AI editing.

01:44 – 02:51

03 · Handing clips to Claude

Opens a new Claude Code session, copies the filename, invokes the video-edit skill. Shows what a raw unedited clip looks like.

02:51 – 03:39

04 · Inside the skill folder

Tour of the video-edit skill structure: assets (logos), knowledge (editing rules and feedback rounds), music library, Remotion templates, SKILL.md.

03:39 – 05:16

05 · Live Claude session

Claude transcribes the clip, fetches real images, reads the plan, selects templates, generates a low-quality render.

05:16 – 06:02

06 · Reviewing the output

Plays the finished short inside the session. Points out head-tracking, real image integration, subscribe button animation, on-screen graphs.

06:02 – 07:13

07 · Rendering pipeline explained

Raw clip, probe, transcribe, auto-correct captions, polish plan, template selection, final fixes, render plus music. Scale argument: 18 shorts, one session.

07:13 – 07:30

08 · Free skill download CTA

Directs viewers to buildloop.ai/prompts/videoeditskill. End screen with 5-step setup instructions.

Atomic Insights

Lines worth screenshotting.

Treating the AI like a new hire you must document every rule for produces better editing results than any prompt engineering trick.
A Remotion template library is what makes AI animations consistent because Claude picks from pre-approved templates rather than inventing effects per clip.
The skill file is the moat: two creators with identical Claude subscriptions get different edit quality based solely on their SKILL.md investment.
Splitting into a low-quality preview render and a full-quality render pass lets you give feedback without the time cost of a full render on every iteration.
Descript handles the two things AI still gets wrong in 2026, detecting genuine retakes and shortening word gaps, so the Claude step starts with clean material.
Head-tracking by probing face position frame-by-frame is a pipeline template, not a per-video custom step.
Fetching real logos and actual photos from the internet inside the Claude session closes the looks-AI-generated gap in finished shorts.
Editing 18 clips in one batch session is the scale argument; human time does not grow linearly once the skill requires no per-video setup.
Every feedback round compounds: updating SKILL.md after each session means the 20th short is edited better than the first at zero additional setup cost.
The knowledge folder inside the skill is the mechanism that makes quality improvements permanent rather than session-by-session.

Takeaway

The skill file is the variable, not the AI

WHAT TO LEARN

What separates a usable AI editing workflow from a broken one is the SKILL.md, a living document you treat as a job description for a new hire.

Generic AI editing prompts fail because they give the model no stable ground; a SKILL.md documenting every rule and past feedback removes that ambiguity session after session.
Splitting editing into a Descript pre-processing step and a Claude production step means each tool handles what it does best rather than asking one tool to do everything poorly.
A Remotion template library constrains AI creativity in the right direction: predefined animation components reduce error rates more than any amount of prompt refinement.
Low-quality preview renders before full renders create a feedback loop without the time cost; treat the preview as a proof stage, not an inefficiency.
Every feedback round invested into updating SKILL.md means the 20th short is edited better than the first, with no additional per-session setup cost.
The scale payoff is in batching: editing 18 clips in one session is only achievable because the skill requires no per-video configuration once it is trained.
Head-tracking and real-image fetching are what close the gap between an obviously AI-generated edit and a publishable short; both are pipeline-level features, not one-off prompts.

Glossary

Terms worth knowing.

SKILL.md: A markdown configuration file at the root of a Claude Code skill that defines tools, workflows, rules, and context the AI should follow. In a video-edit skill it describes every editing decision including animation choices, caption rules, and rendering steps.
Remotion: A React-based library for programmatic video creation. Used here to define reusable animation components that Claude can select and parameterize per clip without generating custom code each time.
Scene export: A Descript export mode that produces one MP4 file per scene or clip rather than a single assembled timeline, enabling downstream per-clip processing pipelines.
Probe step: The first stage of the rendering pipeline where the AI analyzes frame-by-frame face and head position to generate the data needed for dynamic zoom and follow effects, commonly called head-tracking.
Low-quality render pass: A first render at reduced resolution produced for human review and feedback, avoiding the time cost of full-quality rendering until the edit has been approved.
Knowledge folder: A subfolder inside the video-edit skill that stores editing rules, style guidelines, and a running log of every feedback round, the primary mechanism by which edit quality improves over repeated sessions.
Fal.ai Higgs field MCP: An MCP server that exposes AI image generation capabilities inside a Claude Code session, allowing the editor to generate custom visuals without leaving the workflow.

Resources

Things they pointed at.

00:43toolDescript ↗

02:51toolRemotion ↗

02:51toolFal.ai Higgs field MCP ↗

07:13linkVideo edit skill (free download) ↗

Quotables

Lines you could clip.

03:39

“I really acted like, okay, I am going to hire an editor and I am going to have to teach him every single step in editing.”

Reframes a complex technical workflow as an intuitive human analogy, no context needed.→ TikTok hook↗ Tweet quote

07:13

“18 shorts edited in one time. That is not beatable.”

Punchy scale claim with a specific number. Stands alone.→ IG reel cold open↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphor

00:00This is a short from my channel last week and it was fully edited by AI. The captions were written, the b roll was sourced, the music was done and all the animations were created by AI. And, honestly, I spent around ten hours fine tuning a cloth scale to achieve this level of editing because I've seen a lot of YouTubers making videos about AI replacing editors.

00:21But when I looked at the results, I still didn't really like it. So in this video, I will show you exactly what I did to achieve high level edits, what the results look like, and how you can actually start using the skill yourself.

00:33So make sure to stick around until the end for that. I use Descript to record all of my videos and my workflow is not 100% clod.

00:41It starts with Descript. So what I do is I put all of my content in Descript. What I can then do is I can select everything.

00:49I can go to AI tools. And, by the way, this video is not sponsored by Descript. What I do is I click click on shorten word gaps.

00:56I can then make all the gaps smaller. Then I go to a I tools once again and I use remove retakes. Then I go through everything and within a few minutes, I basically have all the clips cut up for me.

01:10What I then do is I do the following. So I go to export and I go to local.

01:17I go to scenes and then all the scenes. So what I will then get is I will get every single short, and by the way, this also works for long forms.

01:26I will probably edit the intro of this video with this exact scale. But we can click on all scenes and we can start exporting it. And I will now show you what that looks like.

01:3518 files is a bit much to now download just for this video and I already posted these shorts, so that doesn't really make sense. So let me now go to folder. Right here, you will see all these scenes.

01:44So what I can then do is I can go to Claude and I can use my video edit scale and say, like, hey, all these videos, one to 18, please edit those for me. So I think the start of this workflow is pretty straightforward.

01:58We need to get our clips cut up and those are raw. So I can now show you exactly what it looks like. This is what we start with.

02:04$400,000,000 earned by two people with AI, and everyone started copying their stack. Telehealth startup.

02:10Two person team, they use Chechiuti, Claude and Grogg for So it is just cut up so that Claude now needs to add animations, zooming effects, following me whenever I say something important and whatever.

02:21So, we have these files and we can now go to cloth. What I do is I open a new session and I can just say the following. So, we copy the name of this file.

02:29We go here and I say, please edit for me. What it then uses is the following skill and I will show you exactly what that skill looks like. We can go to my cloth folder and right here we have video edit.

02:42And as you can see, this is not a simple skill with just a skill dot m d file. It has assets with logos for everything that I'm using. It has knowledge.

02:51So it has all these rules for editing or all the feedback that I ever gave. So the more you use it, the better it gets. We can generate images with the Higgs field MCP.

03:01We have a workflow for long forms, for short forms, but we also have music, as you can see right here. We use Remotion for the actual effects and it has these templates. And these templates are basically all the things that it can choose from to make the actual edit.

03:15And, basically, whenever you're editing something and you think like, hey, I think we should have an animation that looks like this, it will edit as a template. So, once again, every time you use it, it gets better. And then we have the scale dot m d.

03:28So we open that and as you can see, it is very large. I said ten hours, but that is really the minimum I put into this. So I really acted like, okay, I'm going to hire an editor and I'm going to have to teach him every single step in editing.

03:44So what I can now do is I can go to Claude. It will then use the skill. It is asking to access my downloads folder.

03:51That is fine. It is looking at the video dimensions and the duration. It is transcribing the video and we can now look at everything that is happening.

03:57So we have the transcription right here, and it is now really understanding what the script is about and what are the main messages that we need to make clear. As a really good editor really looks at everything that is said and looks like, okay, for the viewer, what is important to highlight?

04:12What do we need to explain better? Very quickly, if you're enjoying this video and learning anything from it, please make sure to check the link in the description where you will learn how to do this like a pro. And also, if you're enjoying this video, make sure to like and subscribe.

04:24And then without any further ado, let's go back into the video. So it is now checking for logos as when we're talking about businesses, I want to show the logo of the actual business. And we are talking about the, uh, Ozempic wave, and it is now looking for a real wave and g o p one pen photos.

04:40So it is also fetching real images from the Internet, basically like a real editor. And honestly, I've never spent so much time into fine tuning a skill. I'm really proud of this one.

04:49And I will show you in a minute why. So it is not creating every single animation on the spot. It is really using templates that I fine tuned already.

04:56And that is why the amount of mistakes is very low as everything is already made in a way that looks good. And here, we have the first version, low quality render of the video. As when you render everything in high quality all the time, it takes way longer.

05:10Right now, we can just look at this, give feedback, and it will make the video better, but also the skill. So let's just look at it. $400,000,000 earned by two people with AI, and everyone started copying their stack.

05:22They used Chechiuti, Claude, and Grok for code and copy, and they created an AI customer service. And this really sounds like the future of solo businesses, and it is.

05:33But here's the part that nobody's looking at, and I don't get why. They didn't get rich because they used AI. They were selling GLP one weight loss And that is something I really like, that we have the actual images in here.

05:46It really shows proactiveness of Claude and also, like, I'm just fine tuning this all the time. So it is getting better and better and that is just awesome.

05:55Like, these images right here, everything it follows me as well. As you can see, my head movements are being followed. So, this is overall already really good.

06:04Like, we have these graphs right here and we have a subscribe button at the end. What we can then do is we can then go back to Claude and we can say, hey, this is good. And if we had some feedback, we could say, like, hey, please adjust this a bit and maybe add an animation like this here.

06:19But that is really how you work with an actual editor as well. Right? You give feedback on something that is created.

06:24But right here, I can now say this is good. Please render in full quality. So what really happens is we have our raw clip.

06:31We then probe, so that is the head movement that is being followed. We then transcribe the whole video and we auto correct words that don't make sense as transcribing can make mistakes and we don't want mistakes in our actual captions. And then we polish with AI.

06:45So we lay out a plan on how to actually achieve the best possible edit. And then we polish it, then we write the plan. So we have templates that it can choose from and that is something you can train over time as well.

06:56Then we do some final fixes, so when the video starts, we have some flare in the screen and everything is just taken care of by these steps. And then we render, we add music and we have our finished video. And really think about the scale you can do this on.

07:10Like, I had 18 shorts and it was all edited in one time. Like, that is not beatable.

07:16And then if you want to get this scale, you should go to buildgroup.ai/prompts/videoeditskill and you can just download the skill for free. So make sure to do that as well.

07:25And if you've enjoyed this video, make sure to like and subscribe and then we'll see each other in the next video.

The Hook

The bait, then the rug-pull.

The opening frame reads AI DID in bold on a split screen alongside the finished short, a proof-before-promise cold open that front-loads credibility. Ten hours of SKILL.md tuning is the price the presenter paid to get there, and the rest of the video is the receipt.

Frameworks

Named ideas worth stealing.

00:43model

Descript plus Claude hybrid editing pipeline

Descript: word-gap trimming, retake removal, scene export
Claude Code: transcribe, fetch images, select templates, probe head position
Low-quality preview render plus human feedback loop
Full-quality render plus music

A two-tool pipeline where each tool handles its comparative advantage: Descript for structural cleanup, Claude for visual production.

Steal forAny creator producing short-form content at volume who wants to reduce manual editing time without sacrificing polish.

03:39concept

New-hire teaching model for AI skills

Frame the SKILL.md as a training document for a new hire editor. Document every rule, every preference, every piece of feedback. The AI improves as the document grows.

Steal forBuilding any Claude Code skill where quality compounds with use, such as customer service scripts, copywriting, or content formatting.

CTA Breakdown

How they asked for the click.

VERBAL ASK

07:13link

“If you want to get this skill, you should go to buildloop.ai/prompts/videoeditskill and you can just download the skill for free.”

End-screen with 5-step setup instructions shown on screen. Clean, specific, low-friction. Mid-video soft CTA at approximately 4:01 points to description link.

MENTIONED ON CAMERA

00:43toolDescript ↗

02:51toolRemotion ↗

02:51toolFal.ai Higgs field MCP ↗

07:13linkVideo edit skill (free download) ↗

Storyboard