When you describe a shot to an AI video model the way you’d describe it to a friend — “get closer”, “show the whole place”, “make it feel dramatic” — the model guesses. Sometimes it guesses well. More often it gives you a generic mid-shot with the camera floating somewhere polite. Film crews never had that problem, because they shared a vocabulary: twenty or so terms that name exactly where the camera is, how wide the frame is, and what the lens is doing. Models like Seedance 2.5 and Veo were trained on an enormous amount of captioned footage, and those captions use film language. When you write “low-angle close-up” instead of “a dramatic shot of her face”, you’re not using jargon to sound clever. You’re speaking the language the model already learned to obey. This guide covers the twenty terms that return the most control per word, grouped the way we group them on a shot list: size, angle, coverage, and lens.
Why Models Obey Film Language
An AI video model doesn’t understand your story. It predicts what pixels usually follow a description. The descriptions attached to its training footage were written by people who label shots the way professionals label shots: “wide shot of a market”, “over-the-shoulder two-shot”, “macro detail of hands”. Broad adjectives like “cinematic” or “epic” appear in captions too, but they attach to everything, so they control almost nothing. A precise term like “insert shot” attaches to a narrow, consistent kind of frame, so the model can reproduce it reliably.
There’s a second reason to learn the vocabulary: it makes your decisions cheaper. In a storyboard-first workflow, every panel carries a shot label before a single credit is spent. If the label says “MS, eye-level, static”, the prompt writes itself and the generation either matches the board or it doesn’t — a pass/fail test you can run in seconds. Without the label, you’re judging generations by mood, and mood judgements drift. The vocabulary turns taste into specification. For the broader system behind this, see the site guide to AI video prompting systems, and for movement terms specifically, camera movement prompts for AI video.
You don’t need fifty terms. Twenty covers nearly every shot a short-form creator will ever board. Learn them in families, because families are how you’ll actually choose under time pressure: first how much of the subject fills the frame, then where the camera sits, then how the shot covers the scene, then what the lens does to space.
Size Terms: ECU To Extreme Wide
Shot size describes how much of the subject the frame holds. It’s the first decision on any board because it decides what the viewer can read: environment, body language, or emotion.
- Extreme wide shot (EWS). The subject is tiny inside a large environment. Use it to place a story: a lone figure on an empty beach road at dawn. Prompt phrase: “extreme wide shot, lone figure small in frame”.
- Wide shot (WS). The full subject plus generous surroundings. The character’s whole body is visible with room to move. This is the shot that sells a location.
- Medium wide shot (MWS). Roughly knees-up. Body language and setting share the frame. Useful for walk-and-talk and product-in-hand shots.
- Medium shot (MS). Waist-up, the default of generated video — which is exactly why you must name it when you want it and name something else when you do not. Faces read, hands read, background still present.
- Medium close-up (MCU). Chest-up. The workhorse for dialogue and reaction. Emotion is legible without losing the shoulders that carry it.
- Close-up (CU). Face fills the frame. Reserve it for the moment that matters; a video that lives in close-up has nowhere to escalate.
- Extreme close-up (ECU). Eyes, a mouth, a single detail of the face. High intensity, short shelf life. One per sequence is usually enough.
Two habits make size terms work. First, pair size with subject: “close-up of the barista’s hands” is a different frame from “close-up of the barista”, and models honour the pairing. Second, respect the ladder. A cut from wide to close-up feels like emphasis; a cut from close-up to extreme close-up feels like alarm. When a sequence feels flat, the sizes are usually all neighbours. Spread them: wide to establish, medium to work, close-up to land. If you generated everything in medium shot — most beginners do — re-boarding with deliberate size changes fixes more than any style word.
Angle Terms: Low, High, Dutch, Overhead
Angle describes the camera’s height and tilt relative to the subject. Models respond well to angle words because captions use them constantly — and because angle changes geometry in ways a model can reproduce.
- Eye-level. Camera at the subject’s eye height. Neutral, honest, invisible. The default for dialogue and documentary-feeling work.
- Low angle. Camera below the subject, looking up. The subject gains size and authority; ceilings and sky enter the frame. Use for heroes, products, and anyone who owns the room.
- High angle. Camera above, looking down. The subject shrinks; vulnerability, smallness, being watched. A child at a big table, a suspect at a desk.
- Overhead / top shot. Camera directly above, frame parallel to the ground. Tables, hands at work, food, maps. AI models do overhead well because the geometry is simple.
- Dutch angle. The camera tilts so the horizon leans. Unease, instability, things going wrong. Strong spice: one Dutch shot in a sequence reads as style; five read as a mistake.
- Ground level. Camera resting near the floor. Feet, wheels, pets, and puddles at eye height. Excellent for scale and for making small subjects monumental.
Angle is where beginners most often fight the model. If you prompt “a powerful CEO” you may get anything; if you prompt “low-angle medium shot of a CEO standing at a boardroom window”, the power is built into the geometry. Angle also interacts with size: a low-angle close-up is intimate and dominant at once, while a high-angle wide makes a person small inside their own story. Name both, every time.
Coverage Terms: OTS, POV, Insert, Two-Shot
Coverage terms describe how a shot functions inside a scene rather than how big it is. These four earn their place on almost every shot list we write.
- Over-the-shoulder (OTS). Camera behind one person’s shoulder, looking at the other. The blurred shoulder in the foreground gives the viewer a body to stand inside. Prompt it fully: “over-the-shoulder shot, out-of-focus shoulder in foreground, speaker in sharp focus”. Models need the foreground detail spelled out or they centre both faces.
- Two-shot. Both subjects in one frame. Shows relationship and distance — who leans in, who turns away. Keep two-shots short in AI video; two moving faces are where identity drift starts.
- POV. The camera is the character’s eyes. Hands entering the frame sell it; floating POV without hands reads as drone footage.
- Insert. A tight shot of an object or action the story needs read clearly: a key turning, a phone lighting up, a ticket torn. Inserts are cheap to generate and save sequences in the edit.
- Reaction shot. A face responding to something off-screen. Often generated separately and cut in; it carries emotion without demanding dialogue.
- Cutaway. A related detail away from the main action — rain on the window during an argument. Buys you edit points and atmosphere.
Coverage is where storyboard thinking pays hardest. A dialogue scene boarded as wide, OTS, OTS, reaction can be generated as four simple, stable shots and assembled in Premiere. The same scene prompted as one long two-person generation will swap faces by the second line. The terms aren’t decoration; they’re a production strategy. Our full planning method is laid out in the storyboard-first workflow.
Lens Terms: Wide, Telephoto, Macro
Lens words tell the model how to render space: how much the background stretches or compresses, and how much of the frame stays sharp. Three terms cover most needs, with focal lengths covered properly in our lens guide.
- Wide-angle lens. More of the room visible, space stretches, edges exaggerate. Energy, small interiors, POV. Faces near the edge distort, so keep faces central.
- Telephoto lens. Background compresses toward the subject and melts out of focus. Flattering for faces, isolating for products. Prompt: “telephoto lens, compressed background, shallow depth of field”.
- Macro. Extreme detail: fabric weave, condensation, a watch mechanism. Models handle macro stillness well; keep motion tiny.
- Shallow depth of field. Subject sharp, background soft. The single most reliable “photographed, not rendered” cue you can add.
- Deep focus. Foreground to background all sharp. For environments, architecture, and any shot where the setting is the information.
Use lens terms to finish a shot description, not to open it. Size and angle decide the frame; the lens decides how the frame feels. And keep depth of field consistent across a sequence — a face that pops against cream-soft bokeh in shot one should not sit in deep focus in shot two unless the story moved rooms. Our AI video QC checklist includes a depth-consistency pass for exactly this reason.
Putting Five Terms To Work In One Prompt
Here’s the vocabulary doing its job. Brief: two friends reunite at a night market stall. Boarded as three shots:
Shot 1 (establish): “Extreme wide shot, overhead angle drifting down, crowded night market at dusk, strings of bulbs, one red food stall in the centre, deep focus, photorealistic.” Size, angle, and focus each carry one decision.
Shot 2 (coverage): “Over-the-shoulder medium close-up, eye-level, out-of-focus shoulder foreground, young woman laughing as she recognises someone off-camera, warm practical bulbs behind her, shallow depth of field.” The OTS term plus the foreground note gives the model a composition, not a wish.
Shot 3 (land): “Close-up, low angle, two hands passing a steaming skewer across the stall counter, telephoto compression, steam catching warm light.” An insert-sized close-up ends the beat on an object, which cuts cleanly to anything.
Notice what’s absent: no “cinematic masterpiece”, no resolution numbers, no emotion adjectives without visual causes. Every term in those prompts names something a camera department would recognise, and that’s deliberate. In my own shot lists I write the size and angle in shorthand first — “MCU, eye-level, OTS” — then build the full prompt around that shorthand. It keeps me honest. If I can’t label the shot in film terms, I probably haven’t decided what the shot is yet, and generating at that point is just paying for the model to decide for me.
A practical drill that works: take one generation you liked and one you threw away, and label both after the fact. Shot size, angle, coverage, lens feel. Nine times out of ten the keeper has a clean label and the reject sits in that mushy middle — medium-ish, eye-level-ish, nothing in the foreground, background neither sharp nor soft. You can fix the next generation in ten seconds once you see it that way, instead of rewriting the whole prompt and hoping. What actually works isn’t memorising twenty definitions in one sitting. Pick five terms you’ll use on every project — for most creators that’s wide, medium close-up, close-up, low angle, and over-the-shoulder — and add the rest only when a shot genuinely demands them. For the frame-by-frame structure these terms plug into, revisit the anatomy of a great AI video prompt. Twenty words of vocabulary will do more for your footage than two hundred words of adjectives ever will.