Claude Code Just Killed Video Editing: Shorts That Post Themselves

You are in the right place. This page is the full guide behind the reel: how I build a YouTube Short that goes out to six platforms without me uploading anything by hand. It still needs me to say go before it renders or posts. Nothing here spends money or ships on its own.

It does not mean unattended. Two gates sit in the middle of the pipeline, and neither one is optional. First, I read the finished script out loud as if I had never seen it before and lock it, because every word I change after the voice is recorded costs a new render. Second, once the video is built and checked, I watch the whole thing on my phone and say go. Nothing renders before the first gate. Nothing posts before the second.

I do not guess at topics. I keep a pool of short-form videos that already beat a bar: over 500 comments. This reel's own topic is a good example of why that bar matters. The video it is built on sits at number 1 out of 257 videos in that pool, with 124,637 comments against a pool median of 1,227. That is not a small edge, it is an outlier.

The remix rule: when I build a hook from someone else's proven opener, I keep at least 70% of the original words and only swap the slots (the tool, the number, the audience). A hook that needs more changes than that gets replaced with a different proven one instead of forced.

The target is 105 to 125 words, plain sentences, one idea per line. Numbers get written as words for the voice engine to read (an AI voice will say a dollar figure digit by digit if you leave it as numerals), and the script keeps a comma-loop shape so the last line runs straight back into the first one when it repeats.

The last line before the loop asks for one thing: a single comment word. Mine is AUTOPOST.

Before anything gets recorded, I have a fresh reader, someone with zero context on the project, read the script cold and flag any line that does not land on the first listen. This script went through two cold reads before it was locked, and both caught real problems: an opening line that broke the loop's grammar, and a payoff line that was too vague to land.

My cloned voice reads the locked script, and an AI avatar performs it on camera, in my own background, so it looks like a normal talking-head take. Every sentence gets one matching visual, picked from a library of effects I keep, 88 of them right now, so nothing repeats twice in a row and nothing sits on screen a beat longer than the line it belongs to.

This is the part the version I remixed does not have. The recipe I started from (an n8n workflow chaining a script model, a voice model and a video model straight to the YouTube upload node) has no check between "rendered" and "posted." Mine does.

Before a video is ever called finished, it runs through two checks. A gaze gate flags any stretch where the avatar is not looking at the camera, because a wandering-eye take reads as broken the moment you notice it. Then a full frame scan pulls stills across the whole video and checks them against a list: text cut off, an empty or black frame, a caption sitting on top of a graphic, something shown before it is said. The video does not count as done until that scan comes back with zero flagged rows. If it does not, I fix it and run the scan again.

Once I say go, the finished Short goes out to Instagram, YouTube Shorts, TikTok, X, Threads and Pinterest, each with its own caption written for that platform. That is six platforms, not seven. Facebook is not a separate post. Instagram has a Recommend on Facebook setting that mirrors the post there on its own, so there is no seventh queue to manage or a duplicate upload to keep in sync.

A few real gotchas from building this, so you do not lose the time I did:

• Pinterest rejects a plain link field. The URL has to go in the description text instead, or the post fails.
• Instagram and Facebook are one linked post, not two. Posting them separately just duplicates the content and splits the comments.
• YouTube defaults synthetic-media Shorts to an AI-labeled disclosure. Decide that on purpose for your own account instead of leaving the default in place without looking at it.

The script is written to read as one continuous sentence when the last line runs back into the first, so a full watch and a rewatch both feel like a clean loop instead of a hard stop. Before I commit to one hook, I write two or three versions of line one and post all of them as separate trial reels on Instagram first. The one with the strongest early hold (lowest skip rate in the first few seconds) plus the most comments becomes the version that goes out everywhere else.

I built this in Claude Code, a coding AI, not a video editor. It reads like a normal conversation: I describe a stage, we build it, I watch the output, and we move to the next one. None of it required me to write the underlying code myself.

If you want to build your own version, the Claude Code Club is where I post the systems as I build them, this one included, and where 8,000+ other builders are doing the same with their own tools.