My Best Short Beat 381 of My Videos. So I Automated Shorts

0:00 This is a movie poster for one of my videos, and this is its trailer. >> 50 normal >> Two shorts with different hooks, a Spanish version, and three thumbnails to

0:09 test. All of it started from one message in Claude Code. The video side is Flova, an AI video agent that Claude Code or Codex can drive through its CLI. I'll show you how it planned each piece, the

0:21 one thing I had to fix, and when you shouldn't use it. Here's my problem. My best short has 149,000 views. Only one of my 382 long videos has more. Yet, most of my videos never

0:33 get a short because launch isn't one clip. It's a trailer, shorts with different hooks, thumbnails to test, other languages. That's a day of work for every upload. So, today I'm testing

0:43 whether one message in Claude Code can use Flova to turn one brief into the whole launch kit. Setup is one paste. Flova site gives you a setup prompt. I pasted it into Claude Code, and it read

0:54 the installer first, then installed the Flova CLI by itself. Then [clears throat] it opened my browser to sign in. One setting first, Flova asks me before it spends credits on video.

1:04 Every launch starts with a brief. Mine is one file, the video's link, the article it's based on, the script, the thumbnail, and a few keyframes, plus a creative direction. Launch it like a

1:14 movie, a heist thriller where the thief is the AI model itself with one hard rule, no people and no faces. My viewers asked for no AI faces on this channel. Then one message. I asked Claude Code to

1:26 have Flova plan every piece, and to show me the plan before anything is generated. It uploads the files through the CLI, and a new Flova project opens with them inside. Here's the split.

1:37 Claude Code starts and manages the run, and Flova's agent organizes the creative project. From that one brief, it wrote a spec for each of the eight pieces, what it's for, its shape, its hook, and where

1:47 it sends viewers. And it explains its choices, like three moving shots per video, so the budget goes into a consistent villain and clear action. The no faces rule made it interesting, a

1:57 heist with no people. Its answer is the weight block, a black block with one orange stripe and no eyes, mouth, or limbs. A separate machine plays the agent shell because the model only asks,

2:08 the shell does the stealing. And for the Spanish version, it didn't pay for new video. It reused short age shots, slowed two of them, so the longer Spanish lines fit, kept the two trigger words

2:18 untranslated, and says the full video is in English. Every video gets a storyboard, shot by shot. The camera move, the narration, and the words on screen before anything is generated. And

2:28 it priced the whole plan before making anything. 2,714 credits piece by piece with 286 left in my budget. Then I approved the plan, and Flova asks before every generation with

2:40 the cost and a running total against my 3,000 credit budget. One 12-second trailer shot, 312 credits. I say yes. Each request starts as a card in Flova. Claude Code shows it to me with the

2:51 cost, then sends my answer back to that card. Before any video, Flova built the look, the villain, the set, then nine still frames, one for each shot. Claude Code checked all nine for faces and

3:02 lettering before it showed them to me. The three trailer shots came the same way, 312 credits each, and the running total kept count against my budget. Then I don't baby sit a browser tab. I ask

3:13 Claude Code for the status, and it reads the real project. What's done, what failed, and the credits used so far. My laptop even lost power mid-run. Flova kept working on its side, and Claude

3:23 Code picked the same run back up without resending anything. And it isn't tied to Claude Code. Codex checks the same project through the same CLI. While it works, here's where everything lives.

3:34 You can also prompt Flova from its home page, and projects keeps every launch in one place. One Flova project holds every piece, every shot, and every file on one canvas. The navigator lists it in order,

3:46 the villain, the vault, and a storyboard for each video, shot by shot. Before planning, the agent reviewed my script, key frames, and thumbnail, and checked its own knowledge bases on screenwriting

3:56 and models. Under the chat box, you choose how much the agent may do alone. Run directly, ask when necessary, or ask before every generation. Each shot is an asset, and so is the frame it starts

4:07 from. Hover a shot for its model, resolution, and length. Open a frame for its prompt and buttons to regenerate or upscale it. And when one asset is wrong, you can tell Flovo what to change in

4:17 plain words right on that asset. And the preview plays any edit with its clips and audio on the timeline below. Here's short A with the hook the agent bet on, two words that flip a model.

4:30 >> Two words. Bonsoir, Elliot. >> [music] >> In a lab demo >> and the same short in Spanish. >> Dos palabras. Bonsoir, Elliot.

4:38 En una prueba de laboratorio >> The rest, the trailer for X and LinkedIn, short B with another hook to test against A, three thumbnails for YouTube's test and compare, and the

4:48 poster for my community tab. Every piece points back to the full video. Honestly, I'm not a great prompter, so not every shot is perfect, but I asked for a heist and the agent built an industrial one,

4:58 machines, gates, and a vault. I hope you like it. One thing wasn't right. In the trailer, the second shot opened on the block alone for about a second before the scene appeared. I don't rebuild

5:08 anything. I asked Flovo's agent in its chat to trim that second from the timeline without generating anything new. Flovo reviews that edit, says it will trim the opening second, and

5:18 generate nothing. And about a minute and a half later, it's done. Before and after. The flash is gone, the trailer is 1 second shorter, and it cost no credits. Each piece is its own edit on

5:29 Flovo's timeline, so a fix like that touches one edit and nothing else. Then Claude Code exports every edit and drops the files into one folder ready to post. On Flovo's side, each export shows up

5:41 under recent downloads. Last step, and the one that saves the most time. I save the whole method as a Flova skill. The pieces, the rules, the heist look, and even today's fix as a rule. In Flova, it

5:53 shows up on the skill page next to Flova's own with a main document and a reference. Share gives a link or a short code. Its share code works in the CLI. So, for next week's video, Codex starts

6:03 the launch from that code, and once the request reach Flova, the full plan came back in about 6 minutes with nothing generated. In Flova, the new project shows up with the four files and a plan

6:14 for nine assets with a cost estimate and nothing spent. Where it's the wrong tool. If you want to compare video models frame-by-frame, go straight to a model. For one quick edit, CapCut or

6:24 Canva is faster. If you need your real screen or real code, record it, and it doesn't publish for you. I review everything first, so who's it for? If you already work in Claude code or Codex

6:35 and you ship content in batches like launches, weekly videos, or campaigns, it's worth trying. This whole kit came to 2,714 credits by Flova's own count under my

6:46 3,000 budget. So, tell me, what would you hand to a pipeline like this, and what would you never let it publish? If you want to turn one brief into a batch of videos with less repetitive work, try

6:56 Flova with Codex or Claude code. Sign up with my link in the description to claim up to 300 free credits. This video was sponsored by Flova.