How to Write a Good Text-to-Video Prompt

Think like a camera operator, not a novelist

A novel can say “she felt the city waking up around her.” A camera cannot film a feeling; it can film a woman opening a café shutter as delivery bikes pass in early light. Models think in pictures because they were trained on pictures, so every abstract word in your prompt is a coin toss. Translate moods into visible things: not “a lonely scene” but “one figure on an empty platform, wide shot, grey morning light”.

This is also why prompts ported from chatbots or image generators often disappoint. Length is not the problem — filmability is. A fifteen-word prompt describing one clear shot will beat a hundred-word prompt describing a plot. If you are new to the whole workflow, our guide on how to make a video from text with AI shows where prompting sits in the full process.

The five parts of a reliable prompt

Reliable prompts contain the same five ingredients. Subject: exactly who or what is in frame, with concrete detail (“a red 1960s scooter”, not “a vehicle”). Action: one verb-led movement (“rides slowly past the camera”). Setting: where and when (“a narrow Lisbon street at golden hour”). Camera: how we watch (“tracking shot at street level”, “static wide shot”, “slow drone push-in”). Light and style: the look (“soft dawn light, documentary style”, “neon night, cinematic”).

Compare two prompts for the same idea. Weak: “An amazing video of a dog running happily in a park, epic and beautiful.” Strong: “A border collie sprinting across wet grass toward the camera, ears back, low tracking shot, overcast morning light, shallow depth of field.” The strong version names the breed, the direction, the ground, the camera and the light — the model has almost nothing left to guess, so it guesses less wrong.

• Subject — concrete nouns, one main subject per clip
• Action — a single, visible movement
• Setting — place, time of day, weather if it matters
• Camera — one movement or “static” if none
• Light and style — the overall look in a few words

One scene, one action

The most common beginner mistake is asking for a story instead of a shot: “A knight enters the castle, fights the guard, and escapes on horseback.” Current models handle a few seconds of continuous action; a three-beat plot guarantees a muddle where nothing reads clearly. Break the story into its shots — the knight approaching the gate; a sword clash close-up; a rider galloping away — generate each as its own clip, and let the edit tell the story.

This rule has a budget benefit too. Short, single-action prompts fail less often, and failed generations still cost credits, as explained in AI video credits explained. Simple prompts are not a lack of ambition; they are how the professionals structure the work, with ambition applied at the storyboard level instead.

Specific beats superlative

Words like “amazing”, “epic”, “stunning” and “high quality” tell the model almost nothing — they appear in the captions of good and bad footage alike. Replace them with the visible facts that create the impression: “epic landscape” becomes “a lone hiker crossing a vast salt flat, extreme wide shot, heat haze on the horizon”. If you cannot picture the adjective, the model cannot either.

Know the known weak spots and write around them. Readable text and logos are unreliable — design shots that do not depend on them, and add real text later in an editor. Crowds, intricate hand actions and complex physics (pouring, tying, juggling) frequently break; a prompt that implies them through a simpler framing (“hands placing a finished knot on the table”) survives where the literal action fails. State the aspect ratio and setting in the tool's controls, not only in words, when the tool offers them.

Prompting people and characters

People are the hardest subject and the most requested one. Give the model a person it can hold onto: an age range, hair, one distinctive item of clothing and a simple expression or action. “A woman in her thirties with curly black hair and a yellow raincoat, laughing as she shakes rain off an umbrella, medium close-up” survives a whole clip far better than “a beautiful woman laughing”. Distinctive details act as anchors; generic beauty is a blur waiting to happen.

Expect faces and hands to hold up better in closer shots than in wide ones, and expect the same character to change between separate generations — a model has no memory of your last clip unless the tool offers a specific character-consistency feature. For a recurring character across a series, image-to-video from one approved still is the reliable route, described in the next section. Real, identifiable people are a different matter entirely: do not prompt for a real person's likeness for commercial work, and check the tool's rules on faces before publishing anything.

When a starting image beats a thousand words

If the look of the first frame matters — a specific product, a consistent character, your own artwork — skip pure text and use image-to-video: supply the still, then prompt only for the motion (“the product rotates slowly as light sweeps across it; camera static”). You have fixed composition, colours and subject in the image, so the model's whole effort goes into movement, and results are markedly more consistent.

Most generators in our image-to-video category work this way, and several text-to-video tools — Kling AI and Runway among them — offer image-to-video inside the same product. A practical hybrid: generate a still with an image tool, approve it, animate it. Two controlled steps beat one uncontrolled one.

Iterate like an editor, not a gambler

When a generation disappoints, resist regenerating the identical prompt hoping for luck, and resist rewriting everything. Name the single biggest problem — wrong camera, morphing subject, action missing — and change only the words responsible for it. Keep a simple log of prompt, settings and result for your best clips; within a week you will have a personal phrasebook of what your chosen model understands.

Build a small library of your proven prompts and remix them: same camera language, new subject; same subject, new setting. This is how creators develop a recognisable style with these tools. When you are ready to choose or change generators, test your library of proven prompts on the free plans in our best free AI video generators roundup — a tool that nails your own prompts is worth more than one with a flashier demo reel.

A prompt makeover, start to finish

Watch the rules combine on one example. First attempt: “A chef cooking delicious food in a kitchen, amazing video.” The result, predictably, is a vague figure stirring something unidentifiable while the camera wanders. The diagnosis writes itself: no specific subject, no single action, no camera decision, and two empty adjectives doing no work. The rewrite: “A chef in a white apron tossing vegetables in a flaming pan, close side view, steam rising through warm kitchen light, shallow depth of field.” One subject, one action, one camera position, concrete light.

If the rewrite still fails, the log tells you where: if the flames misbehave, the action is too complex — drop to “the chef placing a finished dish on a wooden counter, steam rising” and add the drama in the edit with sound. Prompting well is mostly this loop: write filmably, watch honestly, fix the one thing that broke. The complete workflow around the prompt — choosing tools, assembling shots, exporting — is in our step-by-step guide.