AI Video Prompts: How to Write Them With a Free Prompt Sheet

Good AI video prompts read like a line from a shot list, not a wish. Name the subject, what it does, where it happens, how the camera sees it, the light, and the sound, in plain sentences, one shot per prompt. You can do all of it for free with a reusable prompt sheet in Google Sheets: one row per shot, one column per element, and a formula that joins the row into a ready-to-paste prompt. It works well up to a handful of shots. It stops working when a sequence has to hold together across many clips or one character must keep the same face.

What makes AI video prompts work

Google's own prompt guide for its video models breaks a prompt into parts: subject, action, scene or context, camera angle, camera movement, lighting, ambiance, and audio ^[1]. You don't need every part in every prompt. But each part you leave blank is a decision the model makes for you, and models default to the average. A woman walking gives you a stock clip. A seasoned baker carrying a tray of bread through a narrow kitchen at dawn gives you a scene.

The same guide stresses that specificity avoids generic output, and that applies to verbs as much as nouns. Actions carry the motion of the clip: walking, pouring, turning toward the camera, a breeze moving a curtain. Small actions are often more convincing than big ones in a short clip. A blooming flower or a melting city skyline needs time, so Google warns you to keep clip length in mind for anything that unfolds over a longer period.

Vendors publish their own formulas, and they agree more than they differ. Kling's guide uses Subject + Movement + Scene, then optional camera language, lighting and atmosphere. For Kling 4.0 (released to all in October 2026) it adds one sentence first, details after, one camera move per shot, and roughly 50 to 100 words in total. Our Kling AI prompt guide walks through those rules. Google publishes separate guides for Veo and Gemini Omni too, which our Google Flow prompt guide sums up for Flow.

Build a free AI video prompt generator in Google Sheets

A prompt sheet turns guesswork into a repeatable method. Each row is one shot. Each column holds one element. The last column assembles the prompt, so you never forget the light or the sound again. It costs nothing beyond a Google account, and the same layout works in any spreadsheet.

1. Open a new spreadsheet and name the columns in row 1: Shot, Subject, Action, Scene, Camera, Light, Sound, Negatives, Prompt, Result notes.
2. Select the Camera cells, then click Insert > Dropdown. In the Data validation rules panel, under Criteria, enter values such as static shot, slow pan left, dolly in, handheld, aerial. Click Add another item for each one, then Done ^[2].
3. Repeat for Light with values like golden hour, overcast, neon at night, soft window light.
4. In the Prompt column of row 2, type =TEXTJOIN(" ", TRUE, B2:G2). TEXTJOIN joins the cells with a space, and the TRUE argument leaves empty cells out ^[3].
5. Copy the formula down the column. Keep Negatives out of the joined range, because tools handle them differently.
6. After every generation, write one line in Result notes: what worked, what broke, which tool and model.

Dropdowns matter more than they look. Free text drifts into vague words like cinematic or nice light. A short list of named camera moves and lighting setups keeps the vocabulary concrete, and it makes rows easy to compare when one version beats another.

Write the prompt shot by shot

The sheet handles structure. You still have to write each cell well. Work in this order.

1. Subject. Who or what, with two or three visual details. Age, clothing, material, breed, color.
2. Action. One main verb, plus one small detail. If you need two actions in sequence, Runway notes that prompts with several sequential actions may benefit from longer durations.
3. Scene. Where and when. Interior or exterior, time of day, weather.
4. Camera. Shot size first (close-up, medium, wide), then one move. Kling's own rule is one camera move per shot.
5. Light. Name the source and its quality, not a mood word.
6. Sound. Only if the model you will use generates audio.

Here is a mini example of one row, joined.

Close-up of a ceramic mug of black coffee on a wooden cafe counter. Steam rises slowly as a hand slides the mug toward the camera. Early morning, a quiet cafe. Slow dolly in. Soft window light from the left, warm tones. The clink of the mug on wood and quiet cafe chatter in the background.

Read it back as a director would. Could two people picture the same clip from it? If not, a cell is still vague.

Language matters too. Google Flow recommends English and says output quality varies in other languages. Runway says prompts work in any language but are most consistent in English. Higgsfield says its prompts work best in English.

Negatives, sound and dialogue: the mistakes to avoid

Negatives are tool-specific. Google's guide says not to write no walls or don't show walls in a negative prompt; instead, list what you don't want as plain nouns, such as wall, frame ^[1]. Kling's guide for Kling 4.0 goes further and keeps only two negatives: no subtitles and no background music. That is why the Negatives column sits outside the joined prompt. Check each tool's guide and paste only what it accepts.

Sound is the second trap. Google advises stating clearly that you want audio and describing it in separate sentences: sound effects, ambient noise, dialogue. That only helps on a model that makes sound. Luma AI's Ray3.2 (June 2026) output is silent, and Runway's Gen-4.5 has no sound of its own in the web app. Kling AI and the Google Flow models generate audio with the video, so the Sound column pays off there.

Dialogue needs its own format. Google's examples name the speaker, then the line after a colon. Kling asks for dialogue in quotes, written in the target language. Keep spoken lines short. Google DeepMind notes that natural short spoken audio is still an area of active development for its models.

Iterate like a director, not a gambler

The usual mistake is rewriting the whole prompt after a bad result. Change one cell at a time instead. If the motion was wrong, edit Action or Camera and leave the rest alone. Then you learn which element caused the problem, and the Result notes column becomes your own prompt library.

Keep the versions. Google Flow's History panel keeps every earlier version with its prompt, but your sheet works across tools, which matters when you switch models. Retries are not free on paid tools: InVideo AI charges every retry again, and on Runway a result you dislike keeps its cost. A precise first prompt saves retries.

Stop when a row works, and duplicate it for the next shot. Copying the Subject and Scene cells is the cheapest way to keep a sequence looking like one video.

Where hand-written prompts stop working

The sheet method has a ceiling, and it is concrete. One prompt makes one short clip. Veo 3.1 Lite in Google Flow makes 4, 6 or 8 seconds, Runway's Gen-4.5 makes 2 to 10 seconds, and Flow makes no clip longer than 15 seconds per generation. A one-minute video is many separate generations, and you stitch them yourself.

Three things break first:

• Multi-shot sequences. Past five or six shots, keeping cuts, pacing and continuity in a spreadsheet becomes project management.
• A character who must stay the same. Text alone gets re-read on every generation, so faces, clothes and props drift. Luma AI says key elements may change between versions unless the Agent is re-anchored to master reference assets.
• Longer videos with voice, music and captions. The sheet writes prompts. It does not edit.

That is where tools with storyboards, multi-shot generation or project memory take over. A real person's face or voice also needs that person's permission, whatever the tool.

Flow keeps the sheet's habit of one shot per prompt: each row becomes its own clip, and Scenebuilder only puts the clips in order. When the sequence itself has to come out of a single generation, the next tool works differently.

Kling still hands you generated footage, not an edited video. Once the job is a finished piece from a script, with voiceover, music and subtitles and the same characters across generations, an agent that edits takes over and your sheet becomes its brief.

If you want the full method from idea to export, start with how to make AI videos. If you are ready to compare tools that turn a prompt or a script into a finished video, see our guide to text to video AI tools.

From one prompt row to a whole sequence

A structured prompt, written by hand, carries you further than most people expect. Describe one shot at a time: subject, action, scene, one camera move, named light, and sound only where the model makes it. Keep it all in a free prompt sheet with dropdowns and a joining formula, change one cell per retry, and log what worked. When the job grows into a sequence, a recurring character, or a finished video with voice and music, move to a tool that plans shots for you and keep your sheet as the brief.