Text to video AI: write a prompt the generator understands

What an AI video generator actually understands

An AI video generator does not make films. It makes shots: one continuous piece of footage of 5 or 10 seconds, with one place, one or two subjects, one action and one camera movement. If you think in those units, almost every problem with text to video AI disappears before you press Generate.

Think of it like briefing a camera operator for a single take. "A fisherman pulls a net onto a wooden boat at dawn, the camera slowly pushes in" is one take. "A fisherman wakes up, has breakfast, goes to sea, survives a storm and comes home to his family" is a short film, and the model will try to squeeze it into 5 seconds. The result is usually a blurry montage where nothing quite happens.

What these models do well:

• Natural motion: water, smoke, fabric, hair in the wind, crowds walking, leaves.
• Light: golden hour, neon at night, light through a window, fog with backlight.
• Simple camera moves: slow push in, pull back, pan, tracking shot, aerial shot, handheld.
• Clear styles: documentary, product advert, animated cartoon, vintage film look.

What they still do badly, and where you should not spend credits:

• Readable text. Shop signs, slogans, labels and subtitles come out as letter-shaped scribbles. If you need text on screen, generate the clip without it and add the text afterwards in an editor.
• Exact brands and logos. The model does not know your logo and will not reproduce a real one faithfully. For a product, it is more reliable to use a photo of it as the first frame.
• Long stories and continuity. Each clip is generated on its own. A character can change clothes or face from one clip to the next unless you anchor it with the same reference image.
• Hands doing precise work, counting and complex physics. Someone tying shoelaces, a pianist playing a specific chord or three balls bouncing in a set order will often look wrong.

The honest conclusion: if your idea depends on legible text or a precise logo, a text to video generator is the wrong tool for that part. Use it for the moving image and do the graphics somewhere else.

How to write a prompt for an AI video: the scene formula

A good prompt for an AI video answers six questions, roughly in this order. You do not need all six every time, but each missing one is a decision the model will make for you, at random.

1. Place. Where are we? "A narrow street in Lisbon", "a white studio", "a pine forest in winter".
2. Subject. Who or what is the protagonist? Be specific about the few details that matter: "an elderly man in a grey wool coat", "a red ceramic mug".
3. Action. One action, in the present tense. "Pours coffee", "walks towards the camera", "the waves break on the rocks".
4. Type of shot. Close-up, medium shot, wide shot, aerial shot, top-down shot.
5. Camera movement. Static, slow push in, slow pull back, pan left, tracking shot alongside the subject, handheld.
6. Light and mood. Golden hour, overcast and soft, neon at night, candlelight, harsh midday sun.

Put together: "A narrow street in Lisbon after the rain, an elderly man in a grey wool coat walks towards the camera carrying a loaf of bread, medium shot, slow tracking shot backwards, soft evening light reflecting on the wet cobblestones." That is about 40 words, and every word gives the model something to do.

Some bad and good pairs, so you can see the difference:

• Bad: "A beautiful amazing video of nature." Good: "Wide aerial shot over a misty valley at sunrise, a river winding through pine trees, the camera glides slowly forward." The first one is all adjectives and no instructions.
• Bad: "A woman is happy, then sad, then she leaves." Good: "Close-up of a woman reading a letter by a window, her smile slowly fades, static camera, cold morning light." One emotional change fits in 5 seconds; three do not.
• Bad: "Product video for our coffee brand with the logo and the slogan Wake Up Better." Good: "Top-down shot of a cup of black coffee on a wooden table, steam rising, a hand places a croissant beside it, warm morning light, static camera." Add the logo and slogan later.
• Bad: "Cinematic, 8K, ultra-realistic, masterpiece, trending." Good: describe the light and the lens instead: "shallow depth of field, backlit by the setting sun". Quality buzzwords do very little; concrete visual decisions do a lot.

A common piece of advice says that longer prompts are always better. Not quite. Past a certain point, extra details compete with each other and the model drops some of them. If a clip keeps ignoring something, it is usually more effective to remove two other details than to repeat the one it ignores.

For a person or a product that has to look exactly like itself, words are not enough. Use a photo as the first frame (the tool accepts a reference image) or start from animate a photo with AI, where the whole clip grows from your own picture.

Vertical, horizontal or square: choosing the format

Pick the aspect ratio before you write the prompt, not after, because it changes how the scene should be framed. The same description works very differently in a tall frame and in a wide one.

• 9:16 vertical for TikTok, Instagram Reels, YouTube Shorts and stories. The frame is narrow, so favour one subject, centred, and actions that happen up and down or towards the camera: a person walking towards you, a drink being poured, a rocket taking off. Avoid wide landscapes with the interest at the edges: they get lost.
• 16:9 horizontal for YouTube, websites, presentations and TV screens. This is where wide shots, landscapes, pans and two subjects side by side make sense.
• 1:1 square for feed posts and marketplace listings, where the image is shown at the same size on phone and desktop. Good for products and close-ups.

Say it in the prompt too. "Vertical framing, the subject fills the centre of the frame" helps the model compose for 9:16 instead of shrinking a wide composition. And keep platform safe zones in mind: in vertical video the top and bottom are often covered by the app's buttons and captions, so do not ask for anything important at the very edges.

If you already have a long horizontal video and only want vertical clips from it, you do not need to generate anything: the video to shorts maker picks the best moments and reframes them to 9:16 following the face, with word-by-word subtitles.

Test cheap, generate expensive

The most expensive mistake with an AI video generator is spending your best quality setting on a prompt you have never seen moving. Every generation is a roll of the dice: the same text produces a different clip each time, and the first attempt often reveals a problem you did not foresee (the camera goes the wrong way, the subject is too small, the action happens too late).

That is why Airpic offers four quality levels in Text to Video, and why they are designed to be used in sequence:

• Draft: 480p, with sound. Fast and the cheapest option. Use it to check composition, action and camera movement.
• Normal: 768p. A good final version for social media, where the clip will be watched on a phone.
• 2K: the sharpest picture, for large screens or for cropping later.
• Cinema: 720p with sound, the most natural motion and the richest texture. The slowest one.

To give you a real idea of the wait: in our tests on 15 October 2026, a 5-second draft with sound came out in 88 seconds, and the high-quality version at 1280 × 720 px took about 400 seconds. So the draft is roughly four to five times faster. Three or four drafts to get the prompt right take less time than a single final clip, and cost much less.

A workflow that works:

1. Write the prompt with the scene formula and generate a 5-second Draft.
2. Look at what went wrong and change one thing at a time: the shot, the movement or the light. If you change everything at once, you will not know what fixed it.
3. When two drafts in a row show what you want, generate the final version in Normal, 2K or Cinema, depending on where it will be published.
4. Use 10 seconds only when the action really needs it (a slow reveal, a long tracking shot). A 10-second clip costs double and gives the model more time to drift away from your description.

An honest warning: the draft tells you whether the idea works, not exactly what the final clip will look like. The higher-quality models interpret the same prompt with more detail, and small things can change. What carries over well is the composition and the camera movement, which is precisely what drafts are for. For the current cost of each level, check the text to video AI tool page. Credits are refunded automatically if a video fails or the provider rejects the request, but not if you simply do not like the result.

Step by step: create a video from text in Airpic

This is the exact process, with what you see on screen at each step.

1. Open the tool. Go to Text to Video. There is no account and no sign-up form; your credits are stored in your browser as a code.
2. Write the scene in your own language. Up to 800 characters. You do not need to write in English: describe place, subject, action, shot, camera and light the way you would explain it to a person. If a character, a product or a style has to stay fixed, add a reference image; it will be used as the first frame.
3. Choose format, length and quality. 9:16, 16:9 or 1:1; 5 or 10 seconds; Draft, Normal, 2K or Cinema. For the first try, Draft and 5 seconds.
4. Read the prompt the AI writes. Before generating, our AI (Claude, by Anthropic) rewrites your description as a detailed cinematic prompt in English and shows it to you. Read it: it is the best way to learn what the model will receive. If it has misunderstood something (it made the shot aerial when you wanted it at eye level, or it added rain you did not ask for), correct your own text and try again.
5. Generate and wait with the tab open. A progress bar shows the state. Drafts take about a minute and a half; Cinema takes several minutes.
6. Watch, adjust and step up. Change one element at a time. When the draft is right, generate the same prompt in the final quality and download the MP4.

One thing to know about privacy: unlike the image tools, text to video does not run on our GPUs in the EU. Your text, and the reference image if you add one, goes to the video model provider and to Anthropic to write the prompt. Airpic does not keep the video: it is deleted 1 hour after download, 1 day at most. Do not ask for real, identifiable people without their consent.

Joining several clips. Since each generation is one shot, a longer video is several clips edited together. Generate each shot separately (the same reference image if the same character appears, the same light and style words in every prompt so they match), download them and upload them to the AI video editor. Before editing it asks you a few questions with recommended values (format, target length, pace, subtitles, colour, music), proposes the cut, and you adjust it by talking to it: "shorter second clip", "slower fade at the end". The preview plays free in your browser; you only pay when you export the 1080p MP4. If you want word-by-word captions on the final video, the editor can add them, and our guide on adding subtitles to a video for social media explains sizes and safe zones.

If one of the generated clips looks soft next to the others, you can also upscale the video to 4K before editing. When you are ready to write your first scene, open Text to Video and create a video from text with AI, starting with a 5-second draft.

FAQ

Keep reading