When I build small side projects with Claude, the code usually lands faster than the visuals. The landing page needs a hero image, the changelog post needs a cover, and the launch thread needs a 5-second clip. This is the loop I use now: Claude writes and tightens the prompts, and Yeri AI turns them into images and short videos in the browser.

Why split the work this way

Claude is good at structure: it can turn a messy product description into a short, specific visual brief. An image model is good at pixels. Keeping the two jobs separate makes each step easier to check. I ask Claude for the brief, I read it, and only then do I spend generations.

Yeri AI is a web studio for AI image generation and video. It covers text to image, image to image, text to video and image to video, plus an AI image editor, photo effects, templates, a background remover and an upscaler. On the model side you can pick from Seedream, Nano Banana, GPT Image, Kling and Seedance, so the same brief can be tried on more than one model without switching tools.

Step 1: ask Claude for a four-part brief

I paste the feature description and ask Claude for exactly four fields:

• Subject: what is in frame (a laptop on a desk, a phone showing a dashboard, a product box).
• Scene: where it is (pale studio, warm desk, soft interior, street at dusk).
• Light: soft window light, clean highlight, low warm lamp.
• Style: editorial photo, clean product shot, flat illustration.

Then I ask for three variants that change only one field each. This keeps the batch comparable. If every variant changes everything, you learn nothing from the results.

A prompt I reuse for SaaS landing pages:

Subject: a minimal analytics dashboard on a laptop screen
Scene: light wooden desk, small plant, morning
Light: soft window light from the left
Style: clean editorial product photo, shallow depth of field

Step 2: generate the first round as text to image

I paste the brief into the text to image generator and run a small batch. I judge the results on three questions only: can you read the composition at thumbnail size, is the subject complete, and does the colour fit the product's palette. If one fails, I go back to Claude and change one field, not the whole prompt.

Step 3: lock a reference with image to image

Once one image is close, I switch to image to image. The composition stays and I only swap material, background or season. This is much more reliable than trying to describe the same layout again in words. When the frame is right, I run the background remover for transparent product shots and the upscaler for the final hero size. Doing the clean-up last avoids wasting effort on a frame that gets thrown away.

For quick covers, templates and photo effects are a good shortcut. The template gives the layout, I drop in the product image, and the result lines up with the rest of the social grid.

Step 4: only then make the video

When the still is approved, I use it as the first frame for image to video. I ask Claude for a one-line motion note ("slow push in, light shifts from left to right, no camera shake") and keep clips short. With AI video generation a 5-second loop is usually enough for a launch post or a README demo GIF. Kling and Seedance are the two video models I compare for the same still.

Step 5: keep the prompt log

The thing that saves the most time is a plain prompt log. For every image I keep:

1. the four-field brief Claude wrote,
2. the model name,
3. the one change I made between rounds,
4. whether it shipped.

Next time I need a visual for the same product, I start from the log instead of a blank prompt, and the new image matches the old ones.

What this is not

This does not replace a designer for brand work. It replaces the first ten rounds of "what should this even look like" for side projects, changelogs and launch posts, where speed matters more than polish.

If you want to try the same flow, the brief template above works as-is with Claude, and you can run it in Yeri AI for free to see which model fits your product.