TL;DR
A good AI video prompt is not a longer image prompt. It needs three things image prompts never carry: motion language (what moves and how), camera language (how the lens moves), and time language (what happens at second 1 vs second 5). Get those six parts in the right order — subject and action first — keep it to 20-50 words, and your reject rate drops sharply. This guide gives you the structure plus a copy-ready template that works across Veo 3.1, Kling 3.0, Runway Gen-4.5, and the Sora 2 API.
Why video prompts fail when they read like image prompts
Image models render one frozen moment. Video models have to invent every frame in between, so a prompt that only describes the look leaves the model guessing about the motion. The result is a clip where the subject barely moves, the camera floats randomly, and the last second melts.
The fix is to write for time. The 2026 model guides all converge on the same advice: lead with subject and action, name an explicit camera move, and keep the whole prompt short. Kling’s own prompting docs recommend 20-50 precise words and a named camera movement in every prompt; longer prose actively hurts because most video models start ignoring detail past roughly 60-80 words.
Who this is for
Anyone driving a modern text-to-video model in June 2026: Google Veo 3.1, Kling 3.0, Runway Gen-4.5, or the Sora 2 API. The structure is model-agnostic; only the limits and access differ (table below).
The 6 parts of a working video prompt
Write them in this order. Subject and action go first because that is what every current model anchors on.
- Subject + action (the verb matters most). “A red fox walking across a meadow” beats “a red fox in a meadow.” A static noun phrase tells the model nothing should move, so nothing does. One subject, one clear verb.
- Camera movement. Use real cinematography terms the models are trained on:
static shot,slow dolly in,tracking shot following the subject,low aerial drift left,handheld follow,orbit. A named camera move is the single biggest “looks pro vs. looks amateur” lever — every reliably good prompt in the 2026 model guides has one. - Duration + time direction. “5-second clip, single continuous take, subject enters frame at 1s.” Most platforms generate fixed lengths (Veo 3.1 is 8s per generation; Kling 3.0 does 5-10s, longer via Extend), so match your timing cues to the clip you’ll actually get.
- Lighting direction + style. Be specific about three things: direction, temperature, and quality. “Golden hour, warm rim light from camera-left, slight haze, film grain” reads very differently from “overcast, soft diffused light.” Lighting direction shifts mood more than any other single word.
- Motion strength. Most tools expose a motion or dynamics setting. Higher = more change per frame (good for action), lower = stability (good for locked-down product shots). High motion + long clip + complex subject is the classic recipe for drift.
- Composition. “Wide shot, subject lower-right, leading space to the left for it to walk into.” Telling the model where the subject sits and where the empty space is gives the motion somewhere to go.
Then generate and triage: watch for motion drift (movement that wanders off-intent), identity drift (the subject morphing mid-clip), and camera glitches (warps near cuts). Any of those = re-roll. Video has a far higher reject rate than image, so expect to generate several and keep one.
Copy-ready prompt template
Use square-bracket placeholders and swap them per shot:
Subject: [one thing]
Action: [single verb / motion]
Camera: [static / slow dolly in / tracking / orbit / aerial drift]
Duration: [seconds], single continuous take
Lighting: [direction] + [temperature] + [quality/atmosphere]
Motion strength: [low / medium / high]
Composition: [framing] + [subject position] + [breathing room]Filled-in example for a 5-second product b-roll:
Subject: a ceramic coffee cup with rising steam
Action: steam drifting upward, slow
Camera: slow dolly in
Duration: 5 seconds, single continuous take
Lighting: soft window light from camera-left, warm, diffused
Motion strength: low
Composition: medium close-up, cup centered, dark backdrop with negative space top-rightA realistic workflow (5-second b-roll)
This is what actually shipping one clean clip looks like, not the demo fantasy:
- Write the prompt from the template above: subject + action + slow dolly + soft window light + 5s + low motion.
- Generate 3 variations. Expect roughly 1 of 3 to be usable on a clean subject.
- Trim the tail. Models commonly degrade in the final ~0.5-1 second, so generate longer than you need and cut the end.
- For a series of clips, keep the camera language identical across all of them. Matching camera moves make the cuts feel intentional instead of jumpy.
- Need the same subject across multiple clips? Drive from a reference image (image-to-video) instead of text only — text-to-video can’t hold a consistent face or product across separate generations.
Which model for which job (June 2026)
Pick the tool by the shot, not the hype. All figures below are as of June 2026 and change often.
| Model | Best at | Single-clip length | Resolution | Access / entry price |
|---|---|---|---|---|
| Veo 3.1 (Google) | Prompt-following, native synced audio, dialogue | 8s per gen, extend to ~140s | up to 4K | Bundled in Google AI Pro ($19.99/mo); also Gemini API / Vertex AI |
| Kling 3.0 (Kuaishou) | Camera motion, character physics, cheapest commercial use | 5-10s, extend to ~3 min | native 4K | Free tier (66 credits/day); paid from $6.99/mo |
| Runway Gen-4.5 | In-platform editing + fine camera control | short clips, editor-driven | up to 4K | from ~$15/mo |
| Sora 2 (OpenAI) | Physics, camera work | API: 4/8/12s (Pro to 25s) | up to 1080p | API only — Standard ~$0.10/sec (720p) |
One important change to know: OpenAI shut down the consumer Sora app and website on April 26, 2026. The Sora 2 API still works for developers but is scheduled to sunset on September 24, 2026, so if you’re picking a tool to build on, Veo 3.1 and Kling 3.0 are the safer long-term bets. (See the Sora / Veo beginner guide for setup.)
Common mistakes
- Writing a paragraph. Past ~60-80 words most models ignore the tail. Aim for 20-50 precise words.
- Saying “cinematic” without naming a camera move. You’ll get the model’s random “cinematic-looking” default. Name the move.
- High motion strength + long clip + complex subject. That combination is drift soup. Drop motion strength or shorten the clip.
- Treating the first generation as final. Video reject rates are high. Budget for 3+ generations per usable clip.
- Ignoring audio settings on Veo. Veo 3.1 generates synced 48kHz audio from the prompt; if you want silence or specific ambience, say so explicitly.
FAQ
Which AI video tool is best in 2026? It depends on the shot. Veo 3.1 leads on prompt-following and native synced audio; Kling 3.0 is the strongest on camera motion and physics and the cheapest for commercial use ($6.99/mo); Runway Gen-4.5 wins on in-platform editing and camera control. Sora 2 is API-only now. Test two of them on your own style before committing.
Why does my clip glitch or melt at the end? Most models degrade in the final ~0.5-1 second of a generation. Generate a clip longer than you need and trim the tail in editing. Lowering motion strength also helps the final frames hold.
How long can a single AI video clip be? Per generation in June 2026: Veo 3.1 is 8 seconds, Kling 3.0 is 5-10 seconds, the Sora 2 API does 4/8/12s (Pro up to 25s). Longer videos come from chaining or “Extend” — Veo can reach ~140s and Kling up to ~3 minutes by stitching segments.
Can I keep the same character across multiple clips? Not reliably with text alone. Use image-to-video from a fixed reference image, or a model’s character-consistency feature (Kling 3.0 holds characters across cuts better than text-to-video). Keep camera language identical across the set so the cuts read clean.
How many words should a video prompt be? 20-50 precise words. Lead with subject and action, include one named camera move, then lighting, motion, and composition. Longer prose gets truncated by most models.