Text to Video Prompts: Formulas and Examples That Work

ByHank Cai
Creator entering a prompt beside generated video

A good text to video prompt describes what appears on screen and how it moves, in that order, using one subject, one action and one camera instruction. Every major model vendor’s official guidance points the same way: start simple, name the visual and the motion, then add detail only where the first result misses. Below you will find the formula, copy-ready examples for Shorts, podcast promos, product videos and ads, the mistakes that waste credits, and a revision routine.

Techvideoblog
techvideoblog.com
Find a Video Tool That Fits
Compare AI video tools through hands-on reviews, real workflow tests, output quality insights, and verified pricing.

Compare AI video tools

TL;DR

  • Build prompts from five parts: camera, subject, action, setting, and style or lighting. Keep one action per clip.
  • Write for 5 to 10 seconds of footage. Long, busy prompts produce muddled clips.
  • Adapt the same core prompt for each platform by changing the aspect ratio, hook and call to action, not the whole scene.
  • When a result misses, change one variable at a time so you know what caused the difference.
  • Check your model’s own prompting guide, because limits and preferred phrasing differ.

Table of Contents

What makes a text to video prompt effective

An effective prompt reads like a shot description, not a wish. Runway’s help center advises starting with the most important visual and motion components, reducing ambiguity and adding detail as needed. Its guide also says clarity matters more than a rigid formula or a target word count.

Adobe frames it the same way: the prompt is production direction. Control what is shown, how it moves, how it is framed and how it feels, rather than naming a vague concept like “a cool tech video.”

Specificity should also be selective. Pin down the subject and the movement you care about, and leave nonessential details open so the model can fill them in. If you specify the color of every background object, you leave the model no room and often get stiff, odd results.

One more point from my own review work: the prompt is only one input. The model you pick, the clip length and the aspect ratio all change the result. Our AI video generators category lists the tools we compare, and how we test AI tools explains what we look at.

What makes a text to video prompt effective

The core formula: camera, subject, action, setting, style

The vendors publish slightly different formulas, but they converge on the same five ingredients.

Source Formula
Runway (beginner) [Camera] shot of [subject] [action] in [environment], plus supporting details
Google Cloud (Veo 3.1) Cinematography, subject, action, context, style and ambiance
Adobe Firefly Shot type, character, action, location, aesthetic
Alibaba Cloud (Wan) Entity, scene and motion as the basic formula, with aesthetic control and stylization added for advanced prompts

In plain terms, that gives you this working template:

[Camera move and shot size] of [subject with a few identifying details] [one specific action] in [setting]. [Lighting and mood]. [Visual style].

Camera

Google Cloud says cinematography can specify movement, composition, lens and focus, with examples such as dolly, tracking, crane, aerial, pan, POV, wide shot, close-up, low angle, shallow depth of field and wide-angle lens. Alibaba adds that camera choice carries emotion: a push-in suggests intimacy or tension, a pull-out reveals scale or isolation, tracking puts the viewer beside the subject, orbiting stresses importance and a fixed camera signals stillness.

Use that when you want a mood, not just a framing. “Slow push-in on a founder’s face” says something different from “static wide shot of a founder at a desk.”

Camera

Subject and setting

Give the subject two or three identifying details: age range, clothing, material, color. Then give the setting one clear anchor, such as a kitchen at dawn or a rooftop in the rain. Runway’s list of useful details includes subject appearance, environment, lighting, framing and style.

Action and pacing

Adobe recommends specific, dynamic verbs and pacing words like slowly, quickly or gradually. “Walks” is weak. “Walks slowly toward the camera, glancing at her phone” is a direction the model can follow. Runway’s guidance, as summarized in its resources, favors positive phrasing, a single action and clips of roughly 5 to 10 seconds. Say what you want to see, not what you want to avoid.

Lighting and style

Adobe suggests describing lighting, color grade, mood, aesthetic style, weather and time of day. “Golden hour backlight, warm grade, soft film grain” does more work than “beautiful lighting.” If your footage needs lighting repair after generation, our guide to AI video relighting covers when to fix it virtually and when to regenerate.

Text to video prompt examples you can adapt

These are templates we wrote to show the structure. Swap the nouns for your own subject, then run them in your model of choice. Each uses the formula above and stays under one action per clip.

YouTube Shorts and vertical social clips

Vertical 9:16. Handheld close-up of a barista pouring oat milk into an iced latte, the milk swirling slowly through the espresso. Morning window light, warm tones, shallow depth of field. Clean social-video look.

Luma’s advertising guidance for Shorts recommends an immediate visual or verbal hook, the product solving a problem, text overlays, and no horizontal framing or slow reveals. Most models cannot render reliable overlay text, so add captions in your editor. Our guide to video editor auto captions compares tools for that step.

Podcast promo visuals

Text to video will not replace the podcast clip itself, but it can create b-roll to sit behind audio waveforms or quotes.

16:9. Slow dolly-in on a vintage microphone on a wooden desk in a dim studio, dust drifting through a single warm lamp beam. Moody, cinematic, muted teal and amber grade.

For longer interviews, cutting existing footage is faster. Our list of AI tools for content repurposing and the walkthrough on how to use Opus Clip cover that route.

Product video

9:16. Slow orbiting shot around a matte black water bottle on a wet slate surface, droplets rolling down the side. Macro detail on the cap texture. Soft studio light from the left, dark gradient background, premium commercial style.

Product-ad workflows from video-generation vendors suggest putting the hook, setting, product action, camera movement and closing shot in one prompt, then reviewing product details and motion and iterating. In practice, check logos, labels and hands first, because those are where generated product shots usually drift. If you have a real product photo, an image-to-video tool gives you more control: see our image to video AI guide.

Marketing ad with structure

For ads you need more than a pretty shot. Vuela’s published example is a 15-second, 9:16 sneaker ad prompt that specifies shot sequence, color palette, voiceover, timed captions, audience, tone, music, logo and a final shop-now call to action. A simplified version:

15 seconds, 9:16. Shot 1 (0-3s): tracking shot at ground level of a runner’s white sneakers hitting a wet city street at dawn. Shot 2 (3-9s): close-up of the sole flexing, then a low-angle push-in as the runner accelerates. Shot 3 (9-15s): product on a plain background, logo centered. Cool blue palette, upbeat electronic music, energetic tone, for casual runners.

Note the caveat from Alibaba’s documentation: rapid scene changes inside one clip are a known weak spot. If a model struggles, generate each shot as its own clip and join them in your editor. For ads that mimic creator-style footage, see our notes on AI UGC videos and disclosure rules.

B-roll for explainers

Aerial pull-out over a small-town main street at sunset, cars moving slowly, shop lights turning on. Warm golden light, documentary style, wide-angle lens.

Generic establishing shots like this are where text to video earns its keep. We compare the options in our best AI b-roll generators roundup.

Adapting one prompt for different platforms and goals

Once a core prompt works, do not rewrite it for each channel. Ngram’s guidance says to state the job, audience, source of truth, visible format, duration, aspect ratio, required facts and call to action, and to separate mandatory requirements from flexible ones. That last distinction matters: tell the model which things must not change and which are open.

CapUCut’s batch guidance is a useful split. Lock the subject, action, scene, style, camera movement, duration and output format. Vary the audience, platform, hook, call to action, caption density, voiceover tone, pacing and length. CapCut also recommends changing only one or two variables at a time, so you can tell which change caused a difference.

A practical checklist:

  1. Aspect ratio. Say it in the prompt or set it in the tool. Google Ads’ US help page lists horizontal 16:9, vertical 9:16 and square 1:1 video assets for app campaigns, each with a recommended length of 10 to 60 seconds, which is a good reminder to plan all three.
  2. Hook. Vertical short-form needs the point in the first seconds. Pre-roll on YouTube follows the same logic, with Luma recommending a hook inside the first five seconds and a clear call to action in the last three.
  3. Audience. “For busy parents” changes the tone and settings the model chooses. Name the audience once.
  4. Goal. Awareness clips can be atmospheric. Conversion clips need the product, the benefit and the action step visible.

If you are producing many variants from a spreadsheet, test a small sample first. Our guide to batch-creating Shorts from a CSV shows how to catch errors before a full render.

After export, the clip still has to be found. A video SEO guide helps with titles, descriptions and structure so the footage you generated reaches viewers.

Tailor the prompt to the model

The same words behave differently across tools, so read the vendor’s own guide before you settle on a style. A few documented differences:

  • Amazon Nova. AWS says video prompts work best written as image captions or video summaries, not conversational commands, and should include subject, action, environment, lighting, style and camera motion. It states that prompts are generally limited to 512 characters, rising to 4,000 for videos longer than six seconds from one prompt, and suggests placing camera movement at the beginning or end.
  • Runway. Start simple, describe the visual and motion elements clearly, then add detail as needed.
  • Adobe Firefly. Its official documentation says text prompts can define content, emotions, settings, camera angles and camera movements including zoom, pan, tilt, static and handheld motion. See our Adobe Firefly review for how it holds up in a real workflow.
  • Veo and Wan. Both publish structured formulas, covered in the table above. Veo leans on cinematography terms, Wan on entity, scene and motion.

If you juggle several models, our model-aware AI video prompt generator page covers how to keep prompts organized per tool.

Common mistakes that waste generations

Alibaba Cloud’s documentation is the clearest list of what to avoid:

  • Naming specific real people. Describe a character by appearance and role instead.
  • Rapid scene changes in one clip. One clip, one scene. Cut shots together later.
  • Expecting exact text or exact lip-synced dialogue. Add text overlays and voiceover in your editor.
  • Very long or complex action sequences. Break them into separate clips.

Two more habits we see creators fall into, drawn from the guidance above: piling on adjectives instead of verbs, and writing negatives (“no blur, no people”) when positive phrasing gives the model something to aim at.

How to revise when the clip misses

Runway’s advice is to start with a simple prompt, and if a component is missing, reinforce it in natural language. Keeping a consistent prompt structure makes changes easier to track. A routine that works:

  1. Generate once with the five-part prompt.
  2. Write down what is wrong: subject, motion, camera, lighting or style.
  3. Edit only that part of the prompt. Do not rewrite everything.
  4. If your tool supports seeds, AWS’s approach is to isolate wording changes first, then try new seeds to explore variation.
  5. Save the version that worked as your template for the next clip.

The discipline matters more than the wording. Changing six things at once makes every result a mystery.

Our perspective

Most prompt libraries hand you a pile of paragraphs and no way to judge them. We think creators are better served by a repeatable structure, tested against the tool they actually use.

That is how we approach the directory: hands-on use, verified pricing and real workflow tests, with ranking criteria we publish. The same principle applies to prompting. A prompt is only as good as the model, the clip length and the job it does, so we evaluate tools by use case fit, such as Shorts, podcast clips or product ads, rather than by marketing copy.

Our advice is to start from the five-part formula, keep one action per clip, change one variable per revision, and pick the tool whose guidance matches your workflow. If a tool cannot hold your subject steady across a revision, that is a finding worth recording, and it is the kind of thing our reviews are built to surface.

H

Find the right tool for your prompts

A strong prompt still needs a model that fits your format. On TechVideoBlog you can compare AI video tools by use case, from Shorts to ads, with pricing we verify and workflows we test.

Techvideoblog

Browse the full tools directory or start with our best AI video tools picks. For planning a season of short clips, our short-form test plan gives you a step-by-step schedule.

FAQ

How long should a text to video prompt be?

Long enough to cover camera, subject, action, setting and style, and no longer. Runway says clarity matters more than word count. Some models set hard limits: AWS documents a general 512-character cap for Nova prompts, with a higher limit for longer videos.

What is the best prompt structure for beginners?

Use camera, subject, action, setting, then lighting and style. Runway’s beginner version is a shot type of a subject doing an action in an environment, followed by supporting details. Keep one action and add detail only after the first result.

How do I get a prompt to work in vertical format?

State the aspect ratio, such as 9:16, in the prompt or the tool’s settings, and frame for it: close-ups, centered subjects and no wide horizontal reveals. Runway’s examples include vertical 9:16 clips that state the ratio directly in the prompt.

Why does my video not match my prompt?

Usually the prompt asks for too much at once: multiple scenes, complex action or exact text. Simplify to one subject and one action, then reinforce the missing element in plain language. Change one variable per attempt.

Can I reuse the same prompt across models?

You can reuse the core idea, but check each model’s guide. Some prefer caption-style descriptions, some publish their own formulas, and character limits differ. Keep a master prompt and adjust the wording per tool.

Sources

Similar Posts