More Marketing Against The Grain episodes

I Fully Automated my Video Editing With Claude & Cursor thumbnail

I Fully Automated my Video Editing With Claude & Cursor

Published 17 Nov 2026

Duration: 00:32:32

AI models like Opus 5.5 now enable marketers to automate video editing and animation, allowing rapid production of high-quality short-form content through iterative skill refinement and local tool integration.

Episode Description

Free Claude Video Editing Guide: https://clickhubspot.com/kesm Ep. 470 Did you know Claude can edit a short in under 20 minutes? Kipp and guest Moritz...

Overview

The episode explores how AI models, particularly since the release of Opus 5.5 roughly one and a half to two months prior to recording, have become capable of performing the full workflow of a video editor. The process involves three stages: capturing or generating footage, editing and clipping the timeline, and adding animations and after effects. Moritz Krepp demonstrates a practical workflow where a raw phone recording is processed in approximately 15 minutes through a structured skill that directs the AI through cuts, B-roll placement, and animation timing.

The discussion emphasizes that the skill file functions as a detailed task list for the AI, requiring multiple iterations to refine. A feedback loop using a tool called FrameJam allows the creator to annotate specific timestamps with corrections, which the AI then applies in subsequent runs. Supporting tools include FFmpeg for cutting, hyper frames for HTML-to-video conversion, and transcription services like 11 Labs API for accuracy.

The discussion also covers a clips channel experiment where a queue of 60 short-form videos was generated from existing content in roughly 30 minutes of review time, illustrating the velocity-to-quality ratio now possible with AI-assisted editing.

What If

  • What if you replaced your manual video editing with an AI skill that processes a raw phone recording into a finished short in about 15 minutes?

    • Move: Set up a structured skill file in a local desktop app (Cursor or Claude Code desktop) that directs the AI through cuts, B-roll placement, and animation using FFmpeg and hyper frames, then feed it a single raw recording.
    • Why Now?: Opus 5.5 and equivalent models now handle the full editing workflow, and the source notes the capability emerged only "maybe one and a half or two months ago," meaning the tooling is fresh and competition is low.
    • Expected Upside: You could propose going from one video per week to multiple per day; the source describes a queue of 60 shorts generated from existing content in roughly 30 minutes of review, suggesting a solo operator could plausibly multiply output several-fold.
  • What if you built a self-improving feedback file so your AI skill learns your style corrections automatically over time?

    • Move: Instruct your skill to create an additional file called "feedback," append every correction you make after a run, and read that file at the start of each subsequent run.
    • Why Now?: The source describes this as a "living memory" that lets the skill "keep learning as you're using it" without you manually editing the skill file each time, reducing iteration overhead.
    • Expected Upside: Over weeks of use, the skill could converge on your specific pacing, cut style, and animation preferences, potentially reducing per-video feedback passes from two to one or fewer.
  • What if you launched a dedicated clips channel as a low-risk testing ground for hooks and formats before committing to your main content strategy?

    • Move: Create a separate channel, generate multiple short-form variations with different hooks and editing styles from your existing library, and use trial reels or similar programmatic features to gather performance data.
    • Why Now?: The source notes creators are "posting twice or three times a day" and using programmatic variations to test what works, and a decision model can select the best clips from an existing library without new recording.
    • Expected Upside: Data from experiments could inform your main channel strategy; the source describes using a decision model to pick shorts from a library and already seeing one video hit a thousand views on a nascent channel.

Takeaway

  • Run your AI video editing in a local desktop app (Cursor or Claude Code desktop) rather than a web interface, because file management and rendering are more efficient locally.
  • Specify transcription accuracy explicitly in your skill: use the 11 Labs API rather than default Whisper, and add a verification step that cross-checks transcription against video frames to prevent timing and caption misalignment.
  • Create a feedback file within your skill that accumulates all corrections over time so the AI reads its own history on each run and converges on your style without you re-explaining preferences.
  • Direct your agent to specific tools (FFmpeg for cuts, hyper frames for HTML-to-video) rather than letting it choose, since training data may not include newer tools and explicit direction gives you more control over output.
  • Start with a single raw recording and one prompt to see the baseline output, then iterate on the skill file over multiple cycles; the source warns that expecting a perfect one-shot on the first try is "what a lot of people get wrong."

Recent Episodes of Marketing Against The Grain

10 Nov 2026 From AI Babysitter to AI Leader

Define AI playbooks with triggers, inputs, steps, and outputs to automate work reliably without constant supervision.

More Marketing Against The Grain episodes