Upload a reference video to Claude with a prompt. Let it write the generation instructions. Paste them into Google Flow with OmniFlash. Done in under 10 minutes from your phone.
This is the full workflow: three steps, one copy-paste prompt, and no software installation. Suitable for client pitch videos, social content, ads, and any deliverable that needs motion.
Your client booked a call because your marketing looked professional. Your motion graphics need to match. A designer charges $300 to $500 for a 15-second animated title card. After Effects takes months to learn. Most operators skip motion entirely and settle for static images.
This workflow changes the math. You screenshot a reference video you like, feed it to a multimodal AI model, get a complete generation prompt back, and drop that prompt into Google Flow. The whole process takes under 10 minutes and runs on a phone.
The three-step workflow is below. Read all three before you start. Step two contains a prompt you copy verbatim.
Step One: Find Your Reference Video
Your reference is the entire foundation of this workflow. The AI model cannot invent a style you did not show it. The more specific your reference, the sharper the output.
What you need at this step: A video that matches the visual style you want. Not a rough inspiration. An exact visual target.
How to find it: Scroll through your saved content, a competitor's LinkedIn, a brand's Instagram Reels, or a stock footage site like Pexels Video or Coverr. Look for motion graphics specifically: text animations, logo reveals, data visualizations, abstract particle effects, or branded title cards. Screengrab tool reels, conference opener clips, or short product demos all work.
What makes a strong reference:
- Clear lighting, not fast-cut or grainy footage
- Consistent visual style throughout (not a mashup of three different aesthetics)
- Under 60 seconds, ideally 10 to 30 seconds
- The style you want in the final output, not approximately the style
How to save it: Screenshot the video frame, or record a short screen recording of the clip if you need to capture motion specifically. Most phones let you save a video from Instagram or LinkedIn natively. If not, any free screen recorder works.
What good looks like at this step: You have one video saved to your camera roll. You can describe in one sentence what makes it look the way it does: "Dark background, white sans-serif text, slow particle drift, calm pacing."
The mistake that makes this step fail: Picking a reference that is too generic. A "corporate style" or "modern aesthetic" reference tells the AI almost nothing. A clip from a specific SaaS company's product launch tells it everything: the exact fonts, motion speed, color palette, and atmosphere. One specific reference beats five vague ones.
Step Two: Upload to Your AI Model and Get the Generation Prompt
This is where the translation happens. The AI reads your reference video, identifies the visual and motion properties that create the look, and converts those observations into a structured prompt for Google Flow.
Any multimodal AI model that accepts video or image uploads can do this step. Claude handles it well because its vision analysis is detailed and it follows structured output instructions reliably. Gemini, GPT-4o, or any model with vision capabilities that you already use will produce similar results.
What you need at this step: The reference video or screenshot saved in Step One, plus the analysis prompt below.
How to run it: Open your AI model on your phone. Start a new conversation. Upload the reference video or screenshot. Then copy and paste the following prompt exactly:
Copy: Reference Analysis Prompt
Copy this.
You are analyzing a reference video to create a precise video generation prompt for Google Flow. Look at the reference I uploaded. Then produce a structured generation prompt with the following sections: VISUAL STYLE: Describe the color palette (specific hex codes or named colors if identifiable), background treatment, typography style (serif/sans-serif, weight, size relative to frame), and any graphic elements present (particles, shapes, gradients, overlays). MOTION TYPE: Describe the primary motion (text reveals, particle drift, zoom, pan, fade, scale). Note the speed (fast cut, slow drift, medium pace). Note whether motion is looping or directional. MOOD AND ATMOSPHERE: Two to four adjectives that capture the overall feel. Examples: "calm and premium", "energetic and youthful", "clinical and precise". CAMERA AND FRAMING: Note whether the shot is static or moving, wide or tight, centered or offset. FINAL GENERATION PROMPT: Write one dense paragraph that a video generation model can use. Combine all sections above into a single descriptive prompt. Refer to the visual style, motion type, mood, and atmosphere concretely. Do not use vague words like "modern" or "elegant" without attaching specific visual details to them. The prompt should be 100 to 150 words. Output only the four labeled sections followed by the FINAL GENERATION PROMPT. Nothing else.
What good looks like at this step: The model returns a structured response with all five sections. The FINAL GENERATION PROMPT section is a dense, specific paragraph. Words like "slow particle drift against a deep navy background" or "white Helvetica Neue Bold text fading in from 0 to 100 opacity over 1.2 seconds" are good signs. Vague lines like "a modern animated background" are not.
If the output is vague, your reference was probably too generic or the model processed a low-resolution screenshot. Go back to Step One and find a higher-quality reference.
What to copy: Copy only the FINAL GENERATION PROMPT paragraph. That is what goes into Google Flow in Step Three.
The mistake that makes this step fail: Editing the prompt before pasting it into Google Flow. The model produced a prompt tuned to your reference. If you rewrite it to sound more natural or add words that feel right to you, you are introducing style drift. Paste it as-is. If the output is wrong, go back and re-run the analysis with a better reference, rather than patching the prompt manually.
Step Three: Generate in Google Flow with OmniFlash
Google Flow is Google's AI video generation tool. It is accessible from your phone's browser. OmniFlash is the model inside Flow that handles motion graphics and short-form visual content well.
What you need at this step: The FINAL GENERATION PROMPT you copied in Step Two and a Google account.
How to run it:
- Go to labs.google/flow in your phone's browser, or search "Google Flow AI" and open the official Google Labs link.
- Sign in with your Google account.
- In the generation field, paste your FINAL GENERATION PROMPT directly.
- Before generating, find the model selector. It may appear as a dropdown or settings icon near the input field. Select OmniFlash.
- Hit Generate.
The generation takes one to three minutes on most connections. You will see a progress indicator. When it finishes, the clip plays in the browser.
What good looks like at this step: A short clip (usually 3 to 8 seconds) that matches the visual style from your reference. The colors, motion type, and pacing should be recognizable. It will not be identical to the reference. The goal is a clip in the same visual family that you can use directly or as a starting point.
How to save it: Use the download button in the Flow interface. The clip saves to your phone's camera roll as an MP4.
Iterating: Generate two or three variations. On mobile, each generation is a separate run. Keep the prompt the same across runs to see natural variation, or add one specific detail to the prompt between runs to steer the output in a particular direction. For example, if the first run produced a clip that was too fast, add "slow motion, 0.5x normal pace" at the end of the prompt.
The mistake that makes this step fail: Selecting the wrong model. Google Flow offers multiple models and the default may change over time. If your output looks rough or doesn't match the motion style described in the prompt, check whether OmniFlash is actually selected. Other models in Flow produce different results for this type of motion-graphics work.
What You Get Back
A single generation run produces one 3 to 8 second MP4. That clip is ready to drop into:
- A client pitch deck as an opening slide animation
- A social reel as a branded bumper before your talking-head footage
- A product ad as a visual layer over a static background
- An email header or landing page hero section
For longer motion sequences, run multiple generations and stitch them together in any mobile editing app. CapCut, InShot, and DaVinci Resolve (iPad) all handle MP4 imports without quality loss.
What This Does Not Do
This workflow produces short AI-generated motion clips. It does not replace a motion designer for complex, brand-specific animation work. Before you send it to a client, know these limits:
Output length is short. OmniFlash generates clips in the 3 to 8 second range. This is appropriate for intro animations, logo reveals, and short bumpers. For 30-second brand films or animated explainer videos, this workflow is not the right tool.
Free tier has generation limits. Google Flow's free access tier restricts how many generations you can run per day or per month. The limits change as the product evolves. If you hit the cap, you either wait or upgrade to a paid Google One AI plan. Check the current limits at labs.google/flow before committing this to a client deadline.
Quality depends on your reference. A blurry screenshot produces a blurry prompt and a blurry output. A fast-cut reference produces a fast-cut prompt that the model may interpret inconsistently. If your first generation looks nothing like what you wanted, the bottleneck is almost always the reference, not the model.
AI analysis struggles with complex motion. Multi-layer animations with synchronized text, particle systems, and color grading shifts are difficult for the analysis model to fully describe in 150 words. The more layers in the reference, the more the output will simplify.
You will need two or three tries. Budget for three generations on any new style. The first tells you what the model does with your prompt. The second lets you correct one thing. The third lands in the right range most of the time.
If You Only Do One Thing This Week
Pick a piece of content you are already making this week, a post, a deck slide, a product walkthrough. Find one reference video that has the visual style you want for the opening. Run this workflow on it. You do not have to use the output. Run it once to see what ten minutes produces.
Most operators who run it once find two or three places to use it within the same week.
The step-by-step guide and the exact prompt from this article are available to download below. If you want the template without the context, that is the faster route.