Guides & Tutorials · 9 min read · By Mathieu Dugue
I had Claude generate five reels for $15: the cheapest one cost 7 cents
I asked Claude to produce five reels from product photos. The cost tool by tool, the mistakes, and the full skill to download.

When you sell something online, you eventually have to post short videos, and you already know how long each one takes.
Good news if you have photos of your products: they’re enough to put out a clean reel, with music and motion. What almost all of us overlook is that simple route, because we assume an AI-made reel has to involve generated video.
On November 13, I helped a fashion creator with more than 300,000 followers on Instagram. She posts outfit collages, and we wanted to see what they looked like as video. I asked Claude Code to produce five different reels from the same collages, with four sub-agents working at the same time.
I’m giving you everything: the five concepts, what each one cost tool by tool, the mistakes that cost me time, and the full skill to download so your agent can do the same thing on your end, starting from scratch.
Two links in this article are referral links: if you sign up through them, I may earn a commission or credits, and you pay nothing extra.
In short:
- Five reels of 16 to 19 seconds were produced in a single session, for about $15 total.
- The cheapest cost 7 cents and uses no generated images at all: it’s the real product photos, animated.
- Generated video accounts for 71% of the bill, and it distorted at least one garment in every reel it appears in.
- Four agents worked in parallel, each in its own folder, in 36 to 44 minutes.
- None of these reels has been published yet, so I have no view counts to give you.
The full skill, with the scripts and settings, is available to download here: fiveminuteai.com/g/ai-reels. Leave your email and you’ll get it right away. If you get stuck during setup, come ask your question in the community.
1. We started from collages and cut-out photos
This creator already had a small tool for building her collages: a light background, a mood photo, and six or seven pieces laid on top (a coat, a sweater, jeans, boots, a bag, a watch, earrings).
The tool kept each piece as a separate image, on a transparent background, with its exact position. With collages flattened into a single image, all we could have done was zoom in. With the pieces separated, we could bring them in one by one.
So before you dive in, check whether you have cut-out product photos somewhere. If you don’t, start there: that’s what will get you the cheapest reel.
2. The five reels, and what each one cost
| Reel | What you see | AI video | Cost |
|---|---|---|---|
| Animated collage | The pieces arrive one by one to the music | none | about $0.07 |
| Animated magazine | A cover and five spreads, voiceover | none | about $0.78 |
| Ghost outfit | The clothes dress themselves, no model | 5 clips | about $3.41 |
| Narrated AI model | A fictional woman wears the outfit at a restaurant | 4 clips | about $4.16 |
| The creator as the model | The creator in five outfits, from one photo | 7 clips | about $6.61 |
The total comes to about $15. Generated video accounts for $10.71 of that; the rest is split between 35 images (about $3.70, a high estimate) and the music, voices and sound effects (about $0.60, read off the account).
The clips were generated with Seedance 2.5 on Higgsfield. On my plan, $110 gets you 6,000 credits, so a 5-second clip in 720p costs me 64 cents and the same clip in 1080p costs $1.47.
Want to know which concept fits your product before spending a single credit? Describe it in the community and we’ll tell you where to start.
3. The 7-cent reel
The first reel contains no generated images. The mood photo fills the screen with the title, folds back into its place in the collage, then the seven pieces arrive one after another.
Each piece has its own entrance: the coat tumbles in spinning, the jeans rise from the bottom, the boots step in with two strides, the bag swings by its handle, the watch clicks into place in little jolts, and the earrings bounce. The name of the piece and its brand appear when it lands.
The editing is done with HyperFrames, a free tool that turns an animated web page into a video. The 7 cents are the 18-second music track and the four sound effects, generated with ElevenLabs.
Rendering the 16.5 seconds took 33 seconds on my computer. Since the edit reads the positions straight from the collage, the same script works as a base for the other 34 outfits.
4. What generated video did to the clothes
In the three reels with generated video, at least one garment came out different from the real thing. A short, high-collared trench became long and double-breasted. A straight sequin dress flared out from one shot to the next. A baguette bag turned into a square flap bag.
For someone who recommends specific products, that’s a problem: the person who clicks needs to find the same bag. So we ended each reel with an end card made from the real photos, where everything is accurate.
So if your reel is meant to sell a specific item, keep the real photo for the moment you name it, and save the generated video for atmosphere.
The prompt wordings that worked, including for the reel without a model, are in the skill. If a clip gives you a weird result, show it in the community.
5. The four things that cost me time
A connector that wouldn’t connect. I tried to connect the HyperFrames server to Claude, and the authorization page came back with a redirect address error. I gave up: the tool also works from the command line, no account needed, and the whole edit went through there. So if a connector fights you, check whether the tool has a command line before you keep at it.
The starting image rejected. Seedance 2.5 rejected my starting image as long as I stayed in its default mode, which is designed for text only. You have to pick the mode that accepts references. So if your clip starts from an image and the command rejects it, it’s the mode you need to change; your image isn’t the problem.
An image that was too heavy. Three agents saw their command fail with no clear message. The cause was the size of the starting image: a 3.5 MB PNG got stuck, while the same image as a 400 KB JPEG went through. So convert your images to JPEG before sending them.
Music faster than requested. I asked for a track at 96 beats per minute, it came back at 120, and two agents got the same surprise. If I had synced the entrances to the requested tempo, everything would have been off. So measure the tempo of the track you actually get before you put anything on top of it.
6. Four agents at once, without stepping on each other
While the main session was editing the first reel, four sub-agents produced the other four. Each took between 36 and 44 minutes.
Three things kept it from turning into chaos. Each agent had its own folder and was forbidden from writing anywhere else. They all read the same starting document, with the commands already verified and a cap of 300 credits per reel. And a lock kept two agents from requesting music at the same time, because ElevenLabs only generates one track at a time per account.
Each agent had to finish with a report: the file, its length, what it had spent, and what was weak. I then reviewed each reel on a contact sheet of twelve frames before showing it.
The template for the starting document given to the agents is in the skill; you can reuse it. Come tell us in the community how many agents you ran.
7. Putting your own face in the video
The fifth reel shows the creator herself in five outfits, from a single photo of her. It’s the most expensive one, because it took 16 images to keep five: the first ones made her thinner and smoother than in the photo.
Two tweaks helped: adding a tight crop of her face as a second reference image, and spelling out in black and white what had to stay, for example “round face, do not slim.”
Two flaws remain. She has the same smile in all five shots, because we only had one photo. And everything below the shoulders is invented, so only she can say whether it’s accurate.
This reel stays a private draft until she has approved it. So if you do this for a client, get their consent in writing before generating, and remember the “AI-generated content” label the platforms require.
What you need to do the same
For the animated collage, you need an agent that runs commands on your computer (I use Claude Code), HyperFrames, which is free, and an ElevenLabs account for the sound.
For the reels with generated images, add an OpenAI key. For the ones with video, add a Higgsfield account. These two links to ElevenLabs and Higgsfield are the referral links I mentioned at the start.
The skill handles the rest: it checks what’s installed, walks you through each account, announces the cost before spending, produces the reel and reviews it.
Download the skill: fiveminuteai.com/g/ai-reels. All the commands are in it, along with each tool’s pitfalls. If a step trips you up, ask your question in the community.
The questions we asked ourselves
Do you need an MCP server for Higgsfield?
No, you don’t: none was configured, and the command-line tool generated all the clips.
Does it work without Claude Code?
The skill is a folder of text files and scripts, so any other agent that can run commands can follow it. I’ve only tested it with Claude Code, on a Mac.
Why generate the sound before the images?
Because the sound sets the rhythm. With a voice, you cut on the sentences. With music alone, you place each action on a measured beat.
How long does one reel take?
An image takes up to two minutes, a clip two to three minutes, and a complete reel with video took a single agent between 36 and 44 minutes.
Can you do the same thing for something other than fashion?
Yes: the pieces arriving one by one become the ingredients of a dish or the screens of an app. I haven’t done it yet, and I’m telling you that straight.
What I haven’t checked
None of these five reels has been published. So I don’t know which one holds attention best, or whether the 7-cent reel gets as many views as the $6.61 one.
I also haven’t listened to the sound: the levels are measured, and the balance of sound effects against the music still needs to be adjusted by ear.
The question I care about now is simple: for your own product, is the real photo, animated, enough, or do you need someone on screen to make people stop scrolling? Come tell me in the community.

