01How to Use Claude Code to Create Videos: The Big Idea
Learning how to use claude code to create videos starts with one line: you ask for a new faceless video on a topic, and Claude Code does the research, script, voice, scenes, edit and thumbnail. Julian Goldie made this whole episode that way, with no camera and no editing software.
When Claude Code sees that request, it loads a skill. A skill is a folder of instructions that Claude reads before it starts. Julian's is called faceless video.
Inside the skill there are two key parts:
- A profile file that holds his name, his voice and his call to action, so every video sounds like him without him repeating himself.
- A quality bar of rules he wrote after the first version got rejected.
The quality bar includes rules like voice from the very first frame, something new on screen every second, and always showing the real product working, never just icons about it. From there, the system runs through ten steps from research to delivery.
02Step 1 and 2: Research With Sources, Then the Script
Claude sends the topic to a research agent. To save tokens, Julian hands this part to the Codex command line with web search switched on. It reads official sources first: the company blog, the docs, the GitHub repo and the paper.
Then it writes a research file, and every fact gets a link next to it. The rule that matters most is simple: if a fact cannot be traced to an official source, it does not go in the script.
That rule has already saved Julian. On one topic, the research said the thing was not a product you can install. It was a research paper.
So the video changed before a single scene was made.
Step two is the script. Claude writes it in Julian's style:
- The tool's name goes in the first line.
- Short lines, with one idea each.
- Every feature becomes a workflow with the exact prompt read out loud.
- Every number comes from the research file, and nothing is made up.
Then Claude stops and waits. Julian reads the script and approves it. That is the only step where he is needed.
03Step 3 and 4: Cloned Voice With Word Timings and Music
Step three is the voice. Claude runs one script that sends the approved text to a voice service that uses a clone of Julian's voice. It gets back two files:
- The audio.
- A list of every word with its exact start and end time.
That word list is the secret to the whole edit. Every caption, every highlight and every cut lands on a word, so the picture always matches what is being said.
Step four is music. Only the first 30 seconds get a music bed. Julian asked for a sense of wonder, with soft strings and a lift at about 20 seconds.
The service makes two takes, and he picks one.
The music sits quietly under his voice and fades out by 30 seconds. After that, it is just the voice.
This split keeps things simple. The voice carries the message, the word list drives the timing, and the music only sets the mood at the start. You do not have to sync anything by hand, because the timings come straight from the voice file.
04Step 5 and 6: Real Pages, Real Footage and Animated Scenes
Step five gathers assets, and it runs while the voice is being made. First come real pages. Claude opens the app, finds the exact text Julian is going to talk about and saves where it sits on the page.
Later, the camera can zoom right onto that line and draw a ring around it.
Second comes official footage. If the company posted a demo video, Claude downloads it, strips the sound and uses short shots of a few seconds each. Every shot gets a credit on screen.
Step six is the scenes. Every scene is a small web page built with HTML and CSS animations. Claude writes one scene for each spoken beat, usually two to six seconds long, timed straight from the word list.
The house style is called studio:
- Soft glowing backgrounds.
- A serif headline on the left that narrates the message.
- The real app in the center with rounded corners and a big soft shadow.
- Call out cards that point at the exact button being discussed.
To keep tokens down, scene writing is split across two or three helper agents. Each gets a short brief and a cheat sheet, and they adapt scenes that already exist instead of designing from zero.
05Step 7 and 8: Claude Checks Its Own Work, Then Renders
Step seven is Julian's favorite part. Before anything renders, Claude takes one still frame every two seconds and lays them out on a contact sheet. Then it actually looks at them.
It checks for:
- Text that overlaps.
- Spelling mistakes.
- Empty frames.
- Small text that is hard to read on a dark background.
- Zooms that miss the thing being talked about.
It fixes what it finds and checks again. This step catches problems before you spend time on a full render, and it means you are not the one hunting for typos frame by frame.
Step eight is the render. Each scene page gets played frame by frame in a browser and recorded at 60 frames a second in full HD. Then the audio gets mixed, and the voice is set to a standard loudness.
Because every scene is a web page, the render is repeatable. If you change one scene, you can render it again without touching the rest of the video.
06Step 9 and 10: Quality Control, Safety Rules and Delivery
Step nine is quality control on the final file. Claude checks the size, frame rate, bit rate and loudness. It looks for black frames.
Then it makes a contact sheet of the finished video and looks at that too.
Safety rules are built in as well:
- Third party footage stays under 30% of the video.
- That footage is always muted and always credited, with Julian's commentary on top.
- The music is instrumental.
- Because the voice is an AI clone, he ticks the altered or synthetic content box when he uploads.
Step ten is delivery. Claude makes 10 different thumbnails. Then it writes 8 to 10 title options and a plain text description with every source listed.
Everything lands in one folder in his downloads: the video, the thumbnails, the titles and the description. Julian opens it, picks a title and a thumbnail, and that is it. The whole job goes from one typed line to one folder that is ready to upload.
07Tips, Feedback Loops and Who This System Is For
Here is the whole system in one view:
- One line in.
- Research with sources.
- A script you approve.
- A cloned voice with word timings.
- Real pages and real footage.
- Scenes timed to every word.
- Claude checks its own work.
- A 60 frames a second render.
- One folder ready to upload.
The best part is that the system gets better every video. When Julian does not like something, he tells Claude once, and it writes that rule into the skill. This week he said no more games in the demos, because the videos are built for businesses.
Now every video remembers that.
A few tips from the episode:
- Keep your name, voice and call to action in a profile file so you never repeat yourself.
- Only allow facts that trace back to an official source.
- Stay in charge of one step: approving the script.
This setup suits creators and business owners who want to publish videos without filming or editing. You copy what already works and change the profile to your own name and voice.
