Text to Video With Veo 3: Six Mistakes That Ruin a Good Clip

Veo 3 rarely fails at random. When a clip comes back with a camera that drifts past the subject, a hand that melts into a counter, or a spoken line that lands half a second after the mouth stops moving, the cause is usually a gap in the prompt. The model filled that gap for you, and it filled it with its own guess.

The mistakes below are the ones that waste the most renders in text-to-video work with Veo 3, roughly in order of how much of a clip they destroy. Each has a correction that costs a phrase rather than a re-render budget, and each is checkable before you click generate.

Mistake 1: Asking one short clip to tell a whole scene

A single generation is a shot, not a scene. Google DeepMind's own Veo samples each describe one subject, one camera behaviour and one audio bed; the published Veo 3.1 examples from Magic Hour are likewise one eight-second generation each, with a review note that the result is one insert shot rather than a complete sequence.

The failure is easy to write by accident: "She walks into the café, orders a flat white, sits by the window and checks her phone." That is four actions competing for eight seconds. What comes back is usually the middle action, performed vaguely, with the phone appearing before the hand reaches the pocket.

Correct it by choosing the one action that carries the shot. Keep the entrance and the phone for separate generations, then cut them together in an editor. If you cannot decide which action matters most, the shot is not ready to prompt.

Mistake 2: Stacking camera moves that argue with each other

"Slow pan, orbit and push in" reads like cinematography and behaves like three conflicting instructions. Magic Hour's troubleshooting list names this exactly: the camera ignores the plan, and the fix is to choose one camera behaviour, state its direction and pace, and remove the movements that fight it.

Watch for the tell in the output. A clip that half-completes a move, pushing forward but never rising, or orbiting a subject that slides out of frame, is usually a prompt with two moves in it rather than a model that cannot move a camera.

Write one move, one direction, one pace: "The camera makes one slow push in, ending on a medium close-up of the hands." Add a second move only after the first one lands on its own.

Mistake 3: Describing a mood instead of a visible action

Words like cinematic, premium, emotional and dynamic do not map to pixels. They are judgements about the finished shot, and the model has no way to render a judgement. What it can render is a surface, a light direction and a movement.

Higgsfield's mistake list makes the same point from the other side: tie emotion to something the camera can see, such as a hesitant hand on a brake or a lingering look, instead of naming the feeling. Magic Hour's repair is a substitution exercise, where premium becomes brushed metal, a matte surface and controlled rim light, and cinematic becomes a specific camera height and lens feel.

Compare "a premium coffee moment" with "steam rising from a white cup on a scratched oak table, low morning light from the left, the camera fixed, the cup and saucer unchanged." The second version gives the model four things it can actually draw and one thing it must not change.

Mistake 4: Writing exclusions as commands

Most prompt writers end with a line like "no text, no watermark, no extra people." Google's own video generation prompt guide asks for the opposite shape: avoid instructive language such as "no walls" or "don't show walls," and describe the unwanted elements as a list instead, naming the objects you do not want to see.

Two practical consequences follow. First, keep exclusions in the form the model was trained to read, which means naming objects rather than arguing with them. Second, treat any exclusion as a request rather than a guarantee. Magic Hour is blunt about this for captions: an instruction to exclude text remains a request, and the final frames still have to be inspected for stray lettering.

So the line becomes "text, watermark, additional people," and the review step becomes a check of the two or three frames where text usually appears: signage, screens, packaging and any surface facing the camera.

Mistake 5: Treating audio as a caption instead of a track

Veo 3 generates audio natively, with sound effects, ambient noise and dialogue coming out of the same pass as the picture. That is the model's biggest advantage over silent video generators, and the fastest way to waste it is to treat sound as an afterthought appended to a visual description.

Two failure shapes dominate. The first is a spoken line too long for the shot: by the time the sentence finishes, the mouth has stopped moving or the line is clipped. The second is a single undifferentiated audio request where music, room tone, effects and speech are piled into one clause, so the model picks one and drops the rest.

Separate the layers and name the speaker. "Quiet studio room tone, one metallic tap as the key touches the table, no music" is a mix instruction. "She says, in a relaxed voice, that these are ready for the first dinner party" is a performance instruction. Keep the line short enough to fit the duration you actually asked for, and count its words against the seconds rather than against the script.

Mistake 6: Leaving contact points and physics open

Anything the prompt leaves unspecified gets decided by the model. Higgsfield's checklist turns this into a concrete pre-flight pass: name the points of contact, meaning hands, fingers and whatever touches what, along with one named light source and the direction of movement.

That is why hands are the classic casualty. A prompt that says "she picks up the key" leaves the grip, the finger count and the angle unresolved, and the render resolves them somewhere between plausible and wrong. A prompt that says "her thumb and forefinger pinch the key by its bow and lift it clear of the table" leaves much less room.

Physics behaves the same way. Cloth, liquid and hair are where the model improvises most, so state what must stay fixed: the jacket does not move, the liquid stays level in the glass, the watch face and strap remain unchanged throughout. These lines are cheap, and they are the difference between a usable product shot and a reshoot.

Fix the instruction, not the whole prompt

The instinct after a bad clip is to rewrite everything, and that destroys your evidence. Change one instruction per retry and you learn which sentence was responsible.

Keep a reference generation: a locked shot of the same subject, with the same audio, that you know works. When a new attempt fails, compare the two and name the difference. If the failure moves every time you re-run identical text, it is a prompt problem. If it is wrong in exactly the same way on every run, it is a setting problem, and more adjectives will not help.

Before rewriting, run the symptom through this order: subject specificity, action count, camera count, light source, contact points, audio layers, then the invariants. The first item that is genuinely underspecified in your text is the one to fix, and usually it is the only one.

A check that catches most of this before you generate

Read the prompt once and tick each line. Where a line has no answer, the model will supply one.

• One subject, described specifically enough to recognise from the text alone.
• One action, stated with a verb that has a visible result.
• One camera behaviour, with a direction and a pace.
• One named light source, with a direction.
• Every contact point named: hands, feet, objects resting on surfaces.
• Audio split into speech, ambience, effects and music, with the speaker named.
• An invariants line: what must not change during the shot.
• Duration, resolution and aspect ratio set in the interface, not in the sentence.

That last item matters more than it looks. A prompt asking for vertical framing does not change the output canvas the way a vertical setting does, and a landscape render cropped to portrait will cut the action you cared about. Where the tool exposes the control, use the control.

None of this makes Veo 3 predictable in the way a storyboard is predictable. It makes failures legible, which is the part you can work with: a clip that misses one named thing can be fixed with one named change, while a clip that misses everything usually means the prompt asked for everything at once. If you are still deciding which shots are worth the render, the Veo 3 use cases that fit short clips cover that choice, and the Veo 3 video generator is where the corrected prompt goes.