EasyHive
CLAUDE RUNS MY BIZ / 39
THE WORKFLOW

AI video faster
than you can watch it

fal took MiniMax's open H3 video model, trained it further and released it as H3 Max. A 15 second clip at 768p comes back in about 15 seconds, and a 5 second clip in under 3. That changes how you work: you stop writing one perfect prompt and waiting, and start testing ten ideas in the time one used to take. This is the workflow, from your first free clip to a batch of clips made by Claude Code.

Sources: fal.ai model pages and docs for MiniMax H3 Max, September 2026. Check the model page for current prices.

Step 1: make your first clip for free

  1. Open the sandbox

    Go to fal.ai/minimax-h3-max and click to generate. Signed in, you get five free generations a day, up to 15 seconds each. No subscription, no card to start.

  2. Pick the mode

    Text to video when you only have an idea. Image to video when you have a real photo (your product, your storefront, your face) and want it to move. Reference to video when the same character or object has to look the same across several clips.

  3. Use these settings to start

    768p, 5 to 10 seconds, aspect ratio 9:16 for Reels and TikTok, prompt expansion on balanced. Keep "quality" expansion off when you want speed: it can add around 30 seconds to every clip.

WHY THE SPEED CHANGES THE WORK When a clip took minutes, you wrote one long prompt and hoped. When it takes seconds, you work like an editor: make five rough versions, keep the best one, change one thing, run it again. The skill that pays now is judging clips fast, not writing the perfect prompt.
Step 2: write prompts the model follows
These rules come from fal's own guide for H3 Max. They matter more than any adjective.
Step 3: three prompts to copy
Replace everything in [BRACKETS]. Each one is written for a 10 second clip.
01B roll for your business (text to video)
0 to 4s: close up of [WHAT YOU MAKE OR DO, e.g. hands folding fresh dumplings] on [SURFACE], warm morning light. 4 to 7s: the camera pulls back to show [THE SPACE, e.g. a small open kitchen] with [ONE DETAIL]. 7 to 10s: [THE PAYOFF, e.g. a plate slides across the counter toward the camera]. Sound: [SOUND, e.g. soft kitchen noise and a pan sizzling]. No on-screen text.
02Make your real product move (image to video)

Upload a clean photo of the product first.

Keep the [PRODUCT] exactly as in the photo: same shape, same label, same colors. Leave the product untouched. Slow camera orbit around it on [SURFACE], [LIGHT, e.g. soft studio light from the left]. [ONE SMALL ACTION, e.g. a drop of water runs down the bottle]. Sound: [SOUND]. Locked focus on the product. No on-screen text.
03A character that talks (text to video)
[CHARACTER, e.g. a woman in her 30s with short black hair and a green apron] stands in [PLACE], slight camera move toward her. She says, warm and confident: "[YOUR LINE, max about 20 words]" Sound: [BACKGROUND SOUND]. Same outfit through the whole clip. No on-screen text.
Step 4: let Claude Code make the batch
When the sandbox proves an idea works, move it to the API and have Claude Code generate ten versions while you do something else.

  1. Get your fal API key

    In your fal dashboard, open API Keys, create a key and save it in your project's .env file as FAL_KEY=.... The API is pay per second of video, so it is not free like the sandbox. Check the current price on the model page before a big batch.

  2. Paste this into Claude Code

    CLAUDE CODE PROMPTCOPY EVERYTHING IN THIS BOX
    Build me a small video batch generator with fal's MiniMax H3 Max. 1. Read my API key from FAL_KEY in .env. Never print it or commit it. 2. Read prompts from prompts.txt, one clip per block, blocks separated by a line with ---. 3. For each block call the endpoint minimax/h3-max/text-to-video with duration [5 TO 15], resolution 768P, aspect_ratio 9:16 and prompt_expansion_mode balanced. 4. Run up to [3] requests at the same time, and retry a failed one once. 5. Download every video into a folder called clips, named 01.mp4, 02.mp4 and so on, in the same order as prompts.txt. 6. At the end, print a table: clip number, first words of the prompt, seconds it took, and the file path. Before running the whole file, test with only the first prompt and show me the result.

  3. Write prompts.txt with Claude

    Ask: "Write 10 prompts for [MY BUSINESS] following the rules in this guide: timed beats that add up to [DURATION] seconds, one sound sentence, no on-screen text." Then run the generator.

WHAT THE CALL LOOKS LIKE (FROM FAL'S DOCS)
npm install --save @fal-ai/client const result = await fal.subscribe("minimax/h3-max/text-to-video", { input: { prompt: "...", duration: 10, aspect_ratio: "9:16" }, }); console.log(result.data.video.url);
Claude Code writes this for you. It is here so you can recognize it.
What to do in the next hour
You have five free clips a day. Spend them like this.
CLIPWHAT TO MAKEWHY
1Prompt 01 about your business, 5 secondsSee the speed and the quality for yourself
2The same prompt with your beats rewrittenLearn what timed beats change
3Prompt 02 with a real product photoCheck it keeps your product intact
4Prompt 03 with one line of dialogueHear the audio it makes on its own
5Your best one at 15 secondsA full clip ready for CapCut
TURN THE CLIPS INTO ONE REEL Download your best clips, drop them in CapCut in order, cut the first and last half second of each so the joins feel tight, then add your captions and text there. Keep the audio the model made where it fits, or lay your own voice over it.
KNOW THE LIMITS One clip is 15 seconds at most, so longer videos are several clips edited together. A character stays consistent inside one clip; across separate clips, use reference to video with the same images. And a render being fast does not make it right: watch every clip before it goes in a post.
WANT MORE LIKE THIS? Every week we take a new AI tool, test it and turn it into something a business can use the same day. The rest of the free resources are in the Easy Hive community classroom.