4K ad in 5 hours12 minByteDance's Seedance 2.5 now renders native 4K, and a Christmas supermarket heist short shows what one person can ship in five hours. The trick isn't the video model. It's handing pre-production to Claude, wiring it to a generator over MCP, and locking the camera before you prompt a single frame.Read now →

Drop In a Raw, Messy Take and Get Back a Finished Short: The AI Editing Workflow Creators Are Quietly Using
AI can now take a raw talking-head recording full of pauses and retakes and return a captioned, B-rolled, scored short-form video. Here's the full workflow, built from a fresh Claude account: the project, the connectors, the two skills, the real errors you'll hit, and what this means for editors and for everyone's feeds.
Video editing was supposed to be one of the last creative jobs AI would crack. Too visual, too fiddly, too dependent on taste.
That assumption is breaking faster than most people expected.
Creators are now handing an AI agent a raw talking-head clip, full of long pauses, stumbles and "let me say that again," and getting back a finished vertical short: tight cuts, captions, B-roll, background music, sound effects and little animated inserts. No timeline. No manual editing app. Some say AI-edited shorts are now among their best performers, with individual videos reportedly passing a million views.
Below is the complete workflow, built step by step from a brand-new Claude account with no saved skills or memory, plus the errors you'll almost certainly hit and how to fix them. It works with Claude's Opus 5.5 model, and creators report the same approach works in OpenAI's Codex too.
What the finished system does
Here's the loop once everything is set up:
- You paste a link to something you want to make a short about: a blog post, a research page, a product announcement.
- The AI researches it, checks how similar videos performed, and writes you a script with talking points, title ideas, a shot list and a predicted performance score.
- You record yourself talking to camera, using the script as a rough guide, mistakes and all.
- You drop the raw file in. That's the whole prompt.
- The AI edits it: transcribes it, cuts the dead air and retakes, adds B-roll from the source page and your own footage, builds animated inserts, generates music and sound effects, burns in captions, renders and quality-checks the result.
In one real example, a short about Google DeepMind's Genie 4 world model, the agent ran 67 steps on its own: loading the editing skill, cutting the A-roll, preparing B-roll clips and a music bed, writing an edit plan, rendering, reviewing and delivering. It pulled B-roll both from the creator's own test footage and from the official Genie 4 page, generated background music with ElevenLabs, and when there was no good footage of typing a prompt, it built an animated typing sequence itself.
(If you want the background on Genie 4 itself, see our piece on the worlds you can now hear.)
The tool stack, in plain English
You don't need Premiere or DaVinci Resolve. Everything runs "headless," behind the scenes. The pieces:
- Claude (Pro plan or higher) as the brain, with a project to hold instructions and skills to hold your repeatable process.
- vidIQ connector: YouTube research and analytics. It finds outlier videos, pulls transcripts of top performers and helps score ideas. vidIQ offers an official MCP connector for Claude.
- HyperFrames connector: an open-source HTML-based video framework from HeyGen, used here for simple After Effects-style animations, rendered to MP4.
- ElevenLabs connector: transcription with timestamps, background music and sound effects.
- FFmpeg: the free command-line tool that does the actual cutting, layering and rendering.
- Blender: optional 3D objects and animations for extra polish.
- yt-dlp: a script that downloads video from the web for B-roll (more on the legal caution below).
Step 1: Create the project
Start a new Claude project and call it something like "Short-form video editor." In the project instructions, write a plain-English job description. For example:
You are a world-class short-form video editor. You take source material and turn it into a script that is likely to perform well. You clean up A-roll to cut out pauses and mistakes. You edit in B-roll from the source material and from the web to make the video visually appealing. You can use all available tools, including vidIQ, Blender, HyperFrames, FFmpeg and yt-dlp.
That's the agent's standing brief. Every chat inside the project inherits it.
Step 2: Connect the tools
In Claude's Customize area, add the connectors:
- vidIQ: search for it in the connector directory, click connect and authorize your vidIQ account.
- HyperFrames: same process; it sends you to a browser page to confirm.
- ElevenLabs: here's a gotcha. ElevenLabs' docs note that the directory-listed connector omits music. If you want music and sound effects, install the full version (ElevenLabs documents a custom connector URL for this) and sign in.
You'll also need an ElevenLabs API key and a local folder for video files. Add the folder to the project's context and allow access so you don't get prompted every time.
Step 3: Have the AI study what already works
This is the clever part. Instead of describing your style, you let the AI reverse engineer it.
If you already have a channel, prompt something like:
Use vidIQ to analyze my short-form videos on my channel. Look at the best-performing shorts and figure out what makes them perform well. Look at both the script and the visuals. Then create one skill that generates similar scripts that are likely to go viral, and another skill that edits videos to look like my existing videos. The editing skill can use FFmpeg, HyperFrames and Blender. Cut out pauses and mistakes, add B-roll, animations, subtitles, background music and sound effects. When editing, use both the transcript and the waveform.
No channel yet? Name a few channels whose style you admire and ask it to study those instead.
Why "use the waveform" matters
That last line is a pro tip. If the AI cuts only from the transcript, it sometimes chops you off mid-word, because the transcript says a word ended before the sound actually did. Telling it to check the audio waveform makes it cut after the sound ends.
What came back
In the example build, the agent switched to auto mode, ran for about 15 minutes, combed through the channel's outliers (including videos with roughly 3 million, 1.9 million and 1 million views) and produced two skills: a short-script skill and a short-editor skill.
Its analysis was surprisingly specific:
- Topics that win: physical machines doing surprising things, like flying cars, combat robots and body scanners.
- Hook: the series' signature opening line, which it recognized and reused.
- Engagement driver: videos where viewers debate whether the footage is real or AI. (The analysis even flagged that its own visual scan mislabeled some real demo footage as CGI.)
- Visual style: a 50/50 split screen with B-roll on top, magenta captions, and B-roll changing every 1.8 to 3.6 seconds.
Expect this to be iterative. The script skill tends to be close on the first try. The editing skill will probably mess something up. You tell it what to fix, then tell it to update the skill, and repeat until "edit this" is enough.
Step 4: Generate a script from a link
To test the system, the build used MotionBricks, NVIDIA's real-time generative character-animation system, because its project page is full of eye-catching footage.
The whole prompt was the URL. The agent:
- loaded the script skill,
- read the project page,
- searched vidIQ for similar videos,
- scored the visual appeal of the available footage,
- and returned a roughly 50-second script with a 4 out of 5 score.
Its assessment was refreshingly blunt. It said the topic might not stop non-AI viewers mid-scroll, but strong footage existed, there was natural debate potential, and it connected to things people care about (NVIDIA, games, Unreal Engine). It even estimated a view range of 15,000 to 30,000 based on how similar game and 3D shorts had done on the channel.
It also suggested titles and a B-roll shot list, so the edit phase would know what to grab and where. And the script didn't read like typical AI writing; none of the usual "it's not just X, it's Y" tics.
Most creators treat these scripts as guidelines, not something to read word for word. The value is in the talking points, distilled from what's already working.
Step 5: Record, drop, wait
Record the A-roll, ad-libbing freely. The raw take in the example was full of restarts: "Look at it. Walk, jog. Look at it. Walk, jog, and then run…"
Drag the file into the project chat and send it with no prompt at all. The agent recognized it as raw footage, loaded the editor skill, and started: transcribe the rough cut, gather B-roll and visuals, add music, render, review.
The errors you'll hit (and the fixes)
The first full run stopped before finishing. Two problems, both common:
1. Transcription failed. ElevenLabs wouldn't accept that particular video for transcription, and the skill didn't allow anything else to transcribe.
- Fix: tell the agent, "If ElevenLabs isn't working for transcription, use an alternate tool. Update the skill." It rewrote the skill with a fallback.
2. Websites were blocked. Downloads from NVIDIA's site failed.
- Fix: in Claude's settings, under Capabilities, make sure "Create and edit files" is on, turn on "Use the internet while making files," and in allowed websites, choose to allow all sites (or add the ones you need). The agent may ask you to start a new session for the change to apply, and it will hand you the exact prompt to paste.
After those fixes, the edit ran through.
What this can already do (and where it's going)
Beyond cuts and captions, the same setup can:
- generate 3D objects in Blender and drop them into the edit,
- create charts and graphs that appear as you mention numbers,
- add animated overlays on top of B-roll,
- build missing shots from scratch with HyperFrames.
Right now it's best suited to short-form, talking-head formats. Long-form documentary editing, with pacing and story structure across 20 minutes, is a much harder problem.
What a "skill" actually is
The word sounds magical. It isn't. In Claude, a skill is essentially a folder of instructions (and optionally scripts or reference files) that the model loads only when a task calls for it. Anthropic introduced Agent Skills in October 2025 as a way to package repeatable know-how without stuffing it into every conversation.
For this workflow, that matters in three ways:
- Consistency. Your caption color, cut rhythm and hook formula live in the skill, so every edit follows the same rules.
- Memory without bloat. The skill loads when you drop in footage and stays out of the way otherwise.
- Improvement over time. Every time you say "fix this and update the skill," the fix becomes permanent. After a few dozen videos, the skill encodes most of what a human editor would have learned about your channel.
You can open the skill files and read them. That's worth doing: it's the fastest way to see what the AI thinks your style is, and to correct it directly.
A realistic first week
If you want to try this without losing a weekend, here's a sensible ramp:
- Day 1: Create the project, connect the tools, add your folder and API key.
- Day 2: Run the "study my best videos" prompt and read both skills carefully. Edit anything that's wrong.
- Day 3: Do one full test: link, script, record, drop. Expect errors. Fix them and update the skill.
- Days 4 and 5: Make two more shorts. Note every manual fix you make and feed each one back into the skill.
- Days 6 and 7: Compare AI-edited shorts with your usual edits on retention and watch time, not just views.
By the end of the week you'll know whether the system saves you real hours, and exactly where it still needs a human.
Why this matters
For creators
The bottleneck in short-form video was never ideas; it was hours in the edit. If an agent handles the mechanical 80%, one person can publish daily across YouTube Shorts and Instagram without burning out. Several creators report doing exactly that.
For editors
This is the uncomfortable part, and even creators who use these tools say so openly. Short-form editors will likely find work harder to get. Those who adopt AI might handle several times the volume. Others may be squeezed out.
For everyone's feed
The bigger effect may be a flood. When the barrier to a polished short drops to "record a take and drag it in," expect far more videos. The value shifts from editing skill to taste and curation: finding genuinely interesting things and putting them in front of people who wouldn't otherwise see them.
Counterpoints and risks
- Copyright is not optional. Downloading other people's videos with yt-dlp for B-roll can violate copyright and platform terms. Prefer your own footage, official press assets you're allowed to use, or properly licensed stock.
- Platform rules on "inauthentic" content. YouTube's monetization policies allow AI tools, but mass-produced, repetitive or templated content can be ruled ineligible. Original commentary and real editing judgment matter.
- Analytics-driven scripts converge. If everyone optimizes against the same outlier data, feeds get more same-y. Your voice is the differentiator.
- Accuracy. An agent that writes scripts from a web page can get facts wrong. Check claims before you record.
- Cost and limits. Long agent runs consume usage. Rendering, transcription and music generation may carry their own API costs.
- It's not one click yet. Expect several rounds of fixes and skill updates before the workflow is reliable.
Practical takeaways
- Start with two skills: one for scripts, one for edits. Keep improving them instead of re-explaining yourself.
- Let the AI study winners first. Your own top videos are the best style guide.
- Always say "use the transcript and the waveform" for cuts.
- Build in fallbacks (alternate transcription, allowed sites) before your first real job.
- Treat AI scripts as outlines. Your delivery is the part that can't be automated.
- Use your own or licensed B-roll to stay on the right side of copyright and monetization rules.
FAQ
Do I need a paid Claude plan? The workflow was built on a $20-a-month Pro plan. Heavier daily use may need a higher tier.
Can I do this with ChatGPT instead? Creators report the same approach works in OpenAI's Codex with GPT models. The connectors and tools differ slightly.
Will it work for long YouTube videos? It's strongest on short-form talking-head content today. Long-form story editing is much harder for current agents.
Is the AI going to post for me? No. This workflow produces finished files. Posting, and the judgment about what's worth posting, stays with you.
The bottom line
The raw-take-in, finished-short-out workflow is real, and it's surprisingly accessible: one project, a few connectors, two skills and some patience for setup errors.
It won't make anyone interesting. It removes the most tedious part of making a short, which shifts the competition to the parts that were always the hardest to fake: what you choose to talk about, and how you say it. For creators, that's freeing. For editors, it's a warning. For viewers, it's a reason to expect more noise, and to value the people who find the genuinely good stuff.
Sources
- Anthropic: Introducing Agent Skills
- vidIQ: How to use vidIQ in Claude and Claude Code
- vidIQ Help Center: vidIQ MCP
- HyperFrames (HeyGen): MCP server for video
- HyperFrames on GitHub
- ElevenLabs: MCP server documentation
- NVIDIA MotionBricks on GitHub
- FFmpeg documentation
- yt-dlp on GitHub
- YouTube Help: Channel monetization policies





