How to Make a Miniature Food Video With Claude Code and Flow
The short answer: let Claude Code plan the shots and write the prompts, use Google Flow to make one first frame per shot, animate each frame into a clip, then trim and join the clips. A short reel needs only about six shots.
The demo below shows that flow, then plays the finished 16-second caramel apple rose reel with its original sound.
What you are watching: the screens in the first half are a recreation, drawn to explain the steps. They are not a real recording of Claude Code or Google Flow, and the prompts on screen were written for the demo. The pictures and clips come from the finished reel, which is AI-generated. The narration is an AI voice from ElevenLabs. I am an independent affiliate of ElevenLabs and may receive compensation for referrals, but this post has no paid link.
Who this suits
This fits you if you want short, faceless food content and are happy to review every shot. Miniature scenes suit AI video well, because viewers expect them to look slightly unreal.
Skip it if you plan to share a recipe people will cook. AI footage proves nothing about whether a recipe works. Skip it too if your audience follows you for real, handmade miniatures. Passing off generated clips as handwork would break their trust.
The workflow
1. Write one brief
Start with the dish, the setting, the length, and the shape of the video. The reel above started from this kind of brief:
Make a 16-second miniature food reel: a caramel apple rose in a tiny Japanese kitchen. Use Google Flow for pictures and video.
Say "vertical 9:16" here, so it carries into every prompt.
2. Have Claude Code plan the shots
Ask for a numbered shot list. Each shot needs one clear action, and it should follow on from the shot before. This reel has six:
- A wide view of the tiny kitchen
- Butter and apple slices in a copper pan
- Cinnamon and caramel over the apples
- Tweezers shaping the apple rose
- The rose on a plate with a caramel drizzle
- A fork opening the rose
The last shot is the payoff. Ending there also helps a short video loop.
3. Lock the details that repeat
Before writing prompts, have Claude Code describe each recurring item once: the pan, the stove, the tweezers, the hands, the plate. Then paste those exact words into every prompt. Changing even a few words can make an object drift between shots.
4. Make a first frame for each shot
In Flow, click the model name in the prompt box, stay on Image, choose a Nano Banana model, and set 9:16. Paste the prompt for one shot and generate. Keep the best picture, then use your clearest early image as a reference for later shots.
Each first frame should show the moment just before the action. A falling ingredient is easier to animate than a change that has to happen inside the clip.
5. Animate each frame
Switch to Video, then Frames, and drag the picture onto the start frame slot. Paste the motion prompt, set 9:16, and generate. Describe one simple action, and add "no voice" and "no on-screen text" if the model invents them.
Google's help page describes these Frames steps. Labels can change and features vary by region, so check the page if yours look different.
6. Check every clip
Look for extra fingers, tools that melt into food, ingredients that change color, and stray text. Regenerate only the shots that fail. This review is the part you cannot hand off.
7. Trim and join
Trim each clip to its best seconds, then join them. Claude Code can do this with FFmpeg, a free command-line tool, and any video editor can do it by hand. Add sound last. Food videos depend on small sounds, so listen on a phone before you publish.
What this guide does not tell you
This is a workflow guide, not a benchmark. It gives no figures for cost, credits, generation time, or how many tries each shot took, because those were not measured. Check the credit cost Flow shows you, and make one test clip before you run every shot.
Limitations
AI food can be physically wrong. Sauces flow the wrong way, and slices appear from nowhere. Fast cuts hide some of this, but not all of it.
Consistency takes work. Hands, tools, and colors drift from shot to shot, even with references.
Sound is usually the weak spot. Generated audio can be thin or missing, so plan to add your own licensed music and effects.
Do not automate the website. A bot that clicks through Flow breaks whenever the page changes, and it may break Google's terms. If you want full automation, use an official API. Keep a person approving the idea and the final video.
Label it. Say the video is AI-generated, and use each platform's AI label where it is required. Publishing many near-identical videos can also hurt reach.
Options I earn nothing from
The simplest alternative is to skip Claude Code. Write the prompts yourself, run them in Flow, and join the clips in a free editor. For one video, that may be all you need.
Or film the real thing. Miniature cooking kits and a phone give you real sound and real hands, and nothing to disclose.