How to make YouTube Shorts with Claude Code

Claude Code is a coding agent from Anthropic. You describe what you want in plain English, and it writes files, runs commands and fixes its own errors inside a folder on your computer. The Claude Code documentation covers installation and the basics.

It was built for software, but a YouTube Short is a surprisingly good fit. A Short can be described almost entirely in text: a script, a list of sources, a voice file, and graphics that can be drawn with code and rendered to video. When every part of a video is text or code, an agent that reads and writes text and code can make it.

This guide explains what an AI YouTube Shorts workflow in Claude Code looks like and why it works. A brand-new faceless channel built this way reached 517,320 views and 398 subscribers in its first 28 days.

Why Claude Code instead of a video app

Most AI video tools generate a clip from a prompt. You get something back, and if it is wrong, you prompt again. That is fine for a one-off, but a channel needs hundreds of videos that look and sound like they belong together.

Claude Code works differently in three ways that matter for a channel:

It keeps rules in a file. A project can hold a CLAUDE.md file that Claude Code reads at the start of every session. Your channel's voice, visual style, script rules and quality bar live there, so each video starts from the same standard.
It works on real files. Scripts, sources, audio and graphics sit in a folder you own. You can open, change or reuse any of them.
It can check its own work. Because the video is built from code and data, the agent can test it: do the numbers on screen match the sources, does any text overlap, does any screen sit still too long.

The result is closer to a small studio than to a video generator. You set the direction, and Claude Code does the production.

The AI YouTube Shorts workflow, step by step

Here is the shape of the pipeline. Each stage hands a file to the next one.

Ideas. Claude Code proposes a long list of video concepts in your niche. You approve or reject them in one sitting.
Research. For an approved idea, every claim is collected with a link and a quote from the source before any script is written.
Script. A short script of around 160 words is written from those sources, with the question first and the answer last.
Review. A second agent with no memory of the first reads the script and flags jumps in logic, contradictions and claims stronger than their sources.
Voice. The approved script is turned into narration with an AI voice, trimmed and fitted to the length of a Short.
Graphics. Animated maps, charts, timelines and text cards are written in HTML and timed to the exact word in the narration, then rendered as a 9:16 video.
Checks. Before anything costs money and again on the finished file, automatic checks look for wrong numbers, overlaps and long holds.
Package and publish. Title, thumbnail frame and a description with sources, then a scheduled upload.

You stay in the loop at the points where judgement matters: picking ideas and reading scripts. The rest is mechanical, which is exactly the kind of work an agent does well.

Why the script gets a second reviewer

The single biggest quality jump in this workflow comes from separating the writer from the reviewer. An agent that wrote a sentence tends to approve it. A fresh agent that sees only the script and the sources has no such attachment, and it catches the sentence where cause and effect skip a step, or where the script says "always" and the source says "often".

The review needs a limit on how many rounds it can run, otherwise writer and reviewer can trade edits indefinitely. With a cap, the script either passes or comes back to you.

Graphics as code, timed to the voice

The part that surprises most people is that there is no editing timeline. Each scene is a small web page: a map that zooms to a country, a bar that grows to a value, a quote that appears line by line. Claude Code writes those pages and records them to video.

Timing comes from the narration. Once the voice file exists, the timestamp of each spoken word is known, so a number can appear the moment the narrator says it. Changing a fact means changing one line and rendering again, rather than re-cutting a sequence.

This is also why the checks are possible. A program can read the scene files and confirm that every figure on screen appears in the source list.

Publishing on a schedule

A Short is only worth making if it gets published, and the workflow covers that too. Videos are queued and released on a fixed rhythm rather than in bursts, so each upload gets its own window and the numbers are easy to read. YouTube Help's Shorts pages explain the platform side of uploading and Shorts features.

Once one video runs end to end, the same pipeline can run in a batch. You approve fifty to a hundred ideas at once, and the videos are built in parallel while you review scripts as they come in.

What you need to start

A computer and a Claude subscription that includes Claude Code.
A voice source for narration. Our guide on AI voiceover for YouTube explains the options.
A YouTube channel. It can be new, and it can be a faceless YouTube channel from day one.
No programming knowledge and no video editing software.

Where the course goes further

This guide shows the shape of the workflow. The Claude Code YouTube course gives you the parts that make it run: the rules file template for your project, the prompts for ideas, research and scripts, the reviewer prompt with its round cap, the voice settings, the graphics system, the checks that catch errors before render, and the batch setup for publishing at volume.

Start with the free first lesson, which maps the whole pipeline on one page and lists the tools you need. The course includes a 50,000-view guarantee for readers who complete every level and follow the publishing schedule.