I picked up a project that needed three short product videos for a wine and spirits brand, made with AI. I hadn't made AI video before. So before producing anything final, I spent time researching how these tools work, running small tests, and writing down what happened. This post covers the process I ended up with, where I struggled, and what I'd tell someone trying it for the first time.

The setup

I used Claude connected to Higgsfield. Higgsfield generates the video. Claude is where I plan the shot, write the prompt, and run it, so everything stays in one place.

Step 1: Prepare your assets first

Input quality affected the output more than anything else I tested. When the image is weak, the model has to guess, and it fills the gaps with things you don't want, like warped labels or a bottle that changes shape halfway through the clip.

For each product, I start with a high-resolution image, ideally a PNG with a transparent background. A clean cutout tells the model exactly what the product is and where it ends.

If the shot needs more than the product, I prepare that too: a background image, a character, any reference that defines the look. The more of the scene you provide, the less the model has to invent.

Step 2: Write the prompt in parts

I don't write prompts as one long sentence. I split them into parts I can change one at a time. A typical prompt looks like this:

Opal-blue bottle from the reference image, label facing camera. Ocean atmosphere, light sea mist in the background. Macro texture on the glass, small water droplets. Golden hour, warm rim light from behind. Low angle, slow push-in. Calm, premium mood. 5 seconds, 9:16.

That's product, atmosphere, detail, light, camera, mood, duration, and aspect ratio. Splitting it up makes testing easier. When a result is wrong, I can usually tell which part caused it and change only that part.

The atmosphere changes with the product: ocean for something coastal, forest for something earthy. Of all the parts, camera angle made the biggest difference. Same product, same light, different angle, and the shot goes from flat to usable.

Step 3: If you're not from film, describe the feeling

I don't have a film background. I didn't know the right terms for shot types or lighting setups, and guessing at them gave me vague prompts and vague results.

What worked was voice mode. I describe the shot the way I see it in my head ("the bottle sits still, light comes from behind, it feels quiet and expensive") and ask the AI to turn that into a structured prompt with proper camera and lighting terms. Then I review it, fix what's off, and run it.

Talking is faster than typing, and you don't need the vocabulary up front. You pick it up by reading what the AI gives back.

Step 4: Run every prompt 2–4 times

This was the hardest part to accept. The same prompt with the same image does not give the same video. Run it twice and you get two different results.

That means one run tells you very little. A good result might be luck, and a bad one might be bad luck. I run each prompt two to four times and judge the prompt on the whole batch. If most of the batch is usable, the prompt is reliable. If only one is, I keep working on it.

Step 5: Track every run in a sheet

This is the step I'd push hardest. It's optional, but without it you end up running near-identical prompts and forgetting which one worked. Here's the sheet I use:

Two columns matter more than they look.

Duration. AI video has trouble with timing and transitions, and a longer clip gives it more room to break. Logging duration lets you see whether failures line up with clip length.

Credits. Every run costs credits. Running four versions of every prompt adds up fast, and tracking cost per run shows where your budget actually goes.

The rating and notes columns are where the learning happens. After a few batches, patterns show up: which angles hold, which settings keep failing. I repeat steps 2 to 5 until a prompt produces usable results across most of its batch. That's how I got to the three final videos.

If you're starting out

1. Spend time on your images before you write a prompt.
2. Build prompts in parts so you can change one thing at a time.
3. If you don't know film terms, describe the shot out loud and let AI write the technical version.
4. Run each prompt more than once before you judge it.
5. Keep a sheet. It turns trial and error into a process you can repeat.

What I'm still working on

Consistency across a set is the big one. Each prompt can be reliable on its own and still not match the other videos for the same brand. I haven't found a dependable way to lock the look across clips yet.

I'm also trying to decide earlier which prompts deserve more runs and which to drop, so I spend fewer credits on dead ends.

Transitions are still the weakest part. When a clip fails, it's usually at the moment something moves or changes.