Edit Shorts With Claude: Matt Wolfe's Workflow and Premiere Handoff

Claude can help turn a talking-head recording into a short video by coordinating research, transcription, cuts, captions, B-roll and a render. In Matt Wolfe's walkthrough, the actual assembly is done headlessly with FFmpeg, not inside Adobe Premiere Pro. That distinction matters: you can use Premiere for the final creative pass, but the video does not demonstrate Claude opening and editing a Premiere timeline.

Watch Matt Build the Workflow

Source note, 4 November 2026: the video and supplied transcript are the source for Matt's workflow, first-person results and the failed first render. The project prompt below was supplied with this update and closely matches Matt's on-screen brief. Official tool documentation was checked for installation and behavior. We have not recreated his edit or measured its time savings. The Premiere handoff and pilot prompt are our own workflow suggestions.

Matt's Claude Project Instructions

Matt enters this kind of high-level brief into a fresh Claude project at 08:07. The brief below preserves the wording provided for this article so readers can copy it. It is a role and workflow brief, not a complete editing skill or an assurance of performance.

Before using it: the sentence about B-roll "around the web" does not grant reuse rights. Tell Claude to use footage you own, material you have permission to use, or assets licensed for the destination platform. Do not let a downloader turn somebody else's public video into unlicensed B-roll; YouTube's terms also restrict downloading and reuse outside authorized paths. Keep final script and asset approval with a person.

From a Source Link to a First Render

Matt begins with a Google DeepMind article about Genie 4, plus B-roll he recorded himself. He uses vidIQ to inspect relevant videos and his own channel data, then lets Claude draft a script and shot list. He treats the script as talking points rather than reading every word. After recording imperfect A-roll, he gives the file to Claude; his editing skill transcribes it, plans cuts, prepares B-roll, adds music and captions, renders a preview and checks the output.

The second demonstration starts from an NVIDIA MotionBricks source page. Matt creates two skills on a fresh Claude account: one for his short-form scripts and one for editing. The editing skill learns from his own successful shorts, including recurring hooks and pacing. Anthropic's skill documentation describes skills as reusable instructions that can be revised as the process improves. In this case, the skills are useful because they capture decisions Matt would otherwise repeat every time.

A workable first version needs far less automation than the demo. Start with one owned A-roll file, one approved source, a few licensed visual assets and a reviewable rough cut. Make a list of candidate trims before rendering. Keep source files separate from output, and ask for an edit decision list with timecodes so a human can see what changed. A transcript alone is not enough to cut cleanly: Matt explicitly asks the skill to check waveforms too, because earlier cuts clipped words.

What the Five Named Tools Actually Do

They are not interchangeable editors. Matt uses 08:00-09:17 to explain the division of labor: research, optional 3D and motion graphics, assembly, and optional media retrieval. vidIQ and the HyperFrames route he uses are hosted connections; Blender, FFmpeg and yt-dlp run on a computer.

Matt also mentions tools that can download web video. Publicly viewable does not mean reusable. Obtain permission or use assets whose license covers the planned edit. A generated graphic can replace missing B-roll, but it should clarify the story rather than make unrelated footage look like evidence.

Install and Connect the Stack, Step by Step

Start with the smallest useful setup: Claude, your own A-roll and FFmpeg. Add vidIQ for research, HyperFrames for motion and Blender for original 3D only when the story needs them. yt-dlp is optional and should never be the default B-roll source.

1. Set up Claude and the project

Open Claude's desktop or web app, create a project called "Short-form video editor," and paste the project instructions above. Give it one source article, your audience and a folder of your own footage. Confirm which files and tools the session can access before asking it to edit. Matt's project and first connector steps start at 07:15.

2. Connect vidIQ for research

Create or sign into a vidIQ account. In Claude, open Customize → Connectors, find vidIQ in the directory and authorize the intended account. If you need its broader tool set, the official help page describes adding https://mcp.vidiq.com/mcp as a custom connector instead. Enable the connector in the conversation, ask it for a few relevant videos or audience questions, and inspect the returned sources. The official repo contains connection metadata and optional workflows; it is not a local copy of vidIQ's hosted data service.

3. Install FFmpeg for the actual edit

Visit FFmpeg's download page and choose a build for your operating system. On Windows, follow one of the Windows build links from that page, extract the archive, and add its bin directory to your user PATH. On macOS or Linux, use the linked package route for your system. Open a new terminal and run ffmpeg -version and ffprobe -version. Both should print version details. The executable is what Claude's local workflow needs; a similarly named Python package is not a substitute.

4. Add HyperFrames only when you need motion graphics

For the route Matt demonstrates, find HyperFrames by HeyGen in Claude's connector directory, connect it and complete the sign-in flow. Ask for one short title or data animation first, then inspect its render before using it in the main edit. The connector works through your HeyGen account. If you prefer a local code project, follow the HyperFrames quickstart and repo; the CLI requires Node.js 22+ and FFmpeg. The hosted connector and local CLI are separate paths. HeyGen also has a short lesson on editing your own footage.

5. Add Blender for original 3D, not for every clip

Download Blender from blender.org, select the installer for your computer, install it and launch it once to confirm it opens. Build and render a simple scene manually first. Only then give Claude a narrow task, such as drafting a Blender Python script for one diagram or object, and review the script and render before using them. Blender's command-line manual covers background rendering; it is an optional advanced route, not a one-click connection to Claude.

6. Add yt-dlp only for permitted downloads

Read the official installation guide. On Windows, the project recommends its released yt-dlp.exe; put it in a dedicated tools folder and run .\yt-dlp.exe --version from that folder. On macOS and Linux, use the matching official release or package instructions. Keep FFmpeg available too; yt-dlp recommends it for merging and processing. Test only with media you own or have explicit permission to download and reuse. Do not give Claude unrestricted permission to collect third-party video for B-roll.

The First Render Was Not the Finish Line

At 21:02, Matt's fresh setup stalls: the transcription service rejects the video and several source URLs are blocked. He changes the skill to permit a transcription fallback, adjusts access, starts a new session and gets a render. The preview still has awkward voice processing, a missed false start, loose gaps and missing music. He gives specific notes and waits for a second pass, which he judges better.

The practical lesson is to review the actual video, not the agent's completion message. Check the first three seconds, every factual claim, clipped syllables, caption timing, speaker audio, visual relevance, licenses and the ending. Matt widened web access to make his demo work; Anthropic's file-creation guidance says broader network access increases risk. Use approved domains and assets for a production workflow, especially if the project includes unpublished footage or client material.

Where Adobe Premiere Pro Fits

If you already edit in Premiere, keep Claude as a planning and rough-cut assistant, then finish in an editor you control. This is an editorial handoff, not a Premiere integration shown by Matt. Ask Claude for a script, shot list, transcript and proposed cut timecodes. Import your source footage into Premiere and use Text-Based Editing to make a rough sequence from speech, then refine cut timing, framing, color and sound on the timeline.

Generate captions from the final sequence, or import an approved SRT as a caption track. Adobe notes that its Text-Based Editing workflow is distinct from the captions workflow; a transcript is not automatically a finished subtitle file. Finally, use Premiere's export controls to check the vertical frame and delivery file. If you instead import Claude's rendered MP4, treat it as a flattened reference cut: individual edits, layers and timing decisions are harder to change.

Try One Clip With Your Own Footage

Run a single 30-60 second test before building a full channel skill. Compare your normal edit time with the total time spent preparing assets, waiting for renders, reviewing and revising. The useful metric is a publishable video you stand behind, not how many clips an agent can output.

This original prompt is a production brief, not Matt's exact wording or a tested automation. It keeps publication and final creative judgment with the editor.

Video Chapters

The example and research

Build the workflow

Test and revise

Sources and Limits

Checked 4 November 2026. This article distinguishes Matt's sponsored demonstration from the documented capabilities of each tool. We did not independently produce a short with his files or verify his audience metrics.