This Free Claude Code Skill Turns Your Agent Into a Video Editor
A free, open-source Claude Code skill guides coding agents through transcribing, cutting, and polishing raw footage into finished video.

What is the Claude Code video editing skill?
It’s a free, open-source “skill” file that teaches a coding agent like Claude Code how to edit raw video footage into a finished product. Instead of dumping footage into an agent and hoping it figures things out, the skill gives the agent a structured pipeline: transcribe the footage, cut out mistakes and dead air, assemble the clips, add transitions and polish, then export. The creator built it over several months by using his own coding agent as his day-to-day video editor, then codifying every recurring fix and judgment call into a reusable skill file anyone can plug into their own setup.
TL;DR
- The skill walks a coding agent through a five-step pipeline: transcription, cutting, assembly, polishing, and export, with validation built into the process.
- Coding agents can’t natively hear audio, watch video, generate images, or clone a voice, so the skill leans on helper scripts and multimodal models to fill those gaps.
- A local-only version of the pipeline works for free using tools like Whisper for transcription, while a more advanced version routes requests through OpenRouter for better transcription, video understanding, thumbnail generation, and voice fixing.
- The creator reports editing a full-length video with the advanced multimodal pipeline for around $2.71 in API costs.
- The skill includes a content-type table that tells the agent which steps to run depending on whether you’re making a talking-head video, a screen-and-face-cam video, shorts, or a course-style recording.
- A lessons file stores natural-language rules learned from past editing mistakes, and a brand file stores colors and style preferences, both meant to be customized per user.
- The output isn’t meant to be publish-ready without a final human check, but the creator says it reliably handles cuts, transitions, and small continuity details that normally eat up editing time.
How does the editing pipeline actually work?
The pipeline is script-driven rather than purely conversational. You tell the agent to “edit this video” and point it to a folder of raw footage, which can be messy and unsorted rather than pre-trimmed clips. From there, the agent runs through five core stages:
- Transcription. The agent generates a transcript with word-level timestamps. This also helps it identify retakes and repeated lines.
- Cutting. Using the timestamp data, the agent identifies dead air, breaths, stumbles, and repeated words that should be removed.
- Assembly. The surviving clips get stitched together in order.
- Polishing. Transitions, animations, and any requested B-roll or motion graphics get layered in.
- Export and validation. The final cut gets checked, often by re-transcribing the exported video and diffing it word-for-word against the original transcript to confirm the cuts landed in the right places.
Each of these steps relies on a dedicated script rather than asking the language model to “understand” the video directly. The coding agent acts as the orchestrator, calling these scripts, receiving structured feedback (timestamps, diffs, flags), and deciding what to do next. This division of labor is the core design idea: deterministic scripts handle anything that requires actually parsing audio or video data, while the agent handles judgment calls about what to keep, cut, or animate.
Why can’t the AI just watch and edit the video on its own?
A language model, even a capable one, can’t natively hear speech, watch footage, generate images, or clone a voice. Those are four distinct capability gaps the skill has to work around:
- Hearing: solved with speech-to-text for transcription.
- Watching: solved by routing footage through a multimodal model that can process video.
- Generating images: needed for thumbnails and visual assets.
- Voice cloning: needed when fixing flubbed lines or patching audio without a full reshoot.
Without scripts and external models plugged in for these tasks, a coding agent is just guessing at structure from file names and folder layout. The skill’s main contribution is bridging that gap: it tells the agent exactly which tool to call for which gap, and how to interpret what comes back.
What’s the difference between the local and multimodal pipelines?
The base pipeline can run entirely locally at no API cost, using something like Whisper for transcription. This gets you most of the way to a usable edit and costs nothing beyond your existing coding agent subscription.
The more elaborate version swaps in multimodal models accessed through OpenRouter, a model routing platform the creator says he relies on heavily across his workflow. This version handles:
- More accurate transcription (the creator notes a model he calls “11 Labs” style transcription service reached through OpenRouter as more reliable than local Whisper for his purposes).
- Actual video review, letting a multimodal model catch visual issues a transcript alone would miss.
- Thumbnail generation using image models.
- Voice fixing, using voice cloning to patch audio issues without needing to reshoot.

Remy.
The tradeoff is cost. Running the full multimodal pipeline isn’t free, but it’s cheap: the creator cites editing one complete long-form video for roughly $2.71 in total API spend. For anyone pricing this against freelance editing rates or the hours spent manually trimming footage, that’s a meaningfully different cost structure.
How do you customize the skill for your own content?
The skill ships with a table mapping content types (talking-head videos, screen recording plus face cam, animated B-roll inserts, shorts, course-style long talks) to the specific steps needed for each. Every content type is built as a combination of the same underlying steps rather than having its own separate documentation, which keeps the skill file from bloating with redundant instructions.
Two files are explicitly designed for customization:
- A lessons file, where small recurring mistakes and fixes get documented in plain language as they come up, rather than requiring a new script for every minor issue.
- A brand file (in JSON), storing colors and style choices used for animations and graphics.
If your content doesn’t match any existing row in the table, the creator’s recommendation is to treat the skill as a starting point and work with the coding agent to add new steps or content types as needed, rather than expecting the default setup to cover every use case out of the box.
Is this worth setting up?
For anyone producing routine video content (screen recordings, tutorials, talking-head explainers, course material) the appeal is clear: it automates the tedious, repetitive 90% of editing (cutting dead air, removing stumbles, assembling clips) while still letting a human review before publishing. It’s less suited to highly stylized or narrative editing where creative judgment matters more than mechanical cleanup.
The honest caveat from the creator himself: output isn’t guaranteed to be upload-ready. Expect to review the final cut and occasionally nudge the agent on specific details. But for format-driven content where consistency matters more than flash, the time savings compound quickly, especially once the lessons file accumulates enough fixes that the same mistakes stop recurring.
Frequently Asked Questions
What coding agents work with this skill?
The skill is designed for Claude Code and similar coding agents that can execute scripts and orchestrate multi-step workflows, since it depends on the agent calling external scripts and interpreting structured feedback.
Do I need to pay for anything to use it?
The skill itself is free and open source. The local pipeline (local transcription, local cutting logic) can run without any API costs beyond your existing coding agent subscription. The advanced multimodal pipeline (better transcription, video review, thumbnails, voice fixing) requires an OpenRouter API key and incurs small usage costs, cited as around $2.71 for one full video.
What types of video does this work best for?
The creator built it around presentation-style and educational content: talking-head videos, screen recordings with a face cam, shorts, and course-style long-form talks. It’s not positioned as a tool for highly produced or narrative-driven video.
Does it fully replace a human editor?
Not entirely. The creator says output typically needs a final human review and occasional direction, but describes the skill as reliable enough that he no longer plans to hire a video editor for his own channel.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
What makes this different from just asking an AI to “edit my video”?
Without a structured skill, a coding agent can get roughly 90% of the way to a usable edit but struggles with the remaining details, like precise cut timing, transitions, and consistency, which often ends up costing more iteration time than manual editing. The skill codifies those edge cases into scripts and rules so the agent handles them reliably instead of needing constant correction.





